The Semantic SEO Glossary: every term, in plain English

Patent evidence 1 patent

• US 7346839 B2 Information retrieval based on historical dataGoogle, 2003 · Expired 2023 · Inventors: Anurag Acharya, Matt Cutts, Jeffrey Dean, Paul Haahr (+6 more)

  The core Google historical-data patent. Claim 1 and dependent claims 10-38 enumerate historical signals including document age, link age, link growth rate, anchor-text age, user-selection history, and query-based click patterns, the exact "Historical Data" variable Koray cites. Supports the formula's numerator by proving search engines accumulate signals over time and score based on their evolution.

Google Content Warehouse leak 9 attributes

• Historical Data (user clicks / sessions)
• QualityNavboostCrapsCrapsClickSignals.goodClicks .badClicks, .lastLongestClicks, .unsquashedClicks, .impressions, stored per-doc click telemetry that is the raw material of Koray's "Historical Data" input to the formula.
• QualityNavboostCrapsCrapsData.country click signals are sliced by country, so "Historical Data" accumulates per market.
• CompositeDocIndexingInfo.rawNavboost the raw per-URL navboost count aggregated at indexing time.
• Topical Coverage
• QualityAuthorityTopicEmbeddingsVersionedItem.siteFocusScore "Number denoting how much a site is focused on one topic." Direct measure of topical concentration.
• QualityAuthorityTopicEmbeddingsVersionedItem.siteRadius "The measure of how far page_embeddings deviate from the site_embedding." Measures breadth around the central topic.
• QualityAuthorityTopicEmbeddingsVersionedItem.pageEmbedding and .siteEmbedding, the embedding vectors coverage is computed against.
• Cost of Retrieval (host efficiency)
• PerDocData.hostAge host freshness window used in Twiddler, lower host age raises retrieval cost through sandboxing.
• IndexingDocjoinerDataVersion.spamContentSiteAgeStats site-age statistics inspected during spam filtering.
• CrawlerChangerateUrlChange.simhashIsTrusted determines whether crawl deduplication can short-circuit, reducing retrieval cost.

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Patent evidence 2 patents

• US 11361227 B2 Onboarding of entity dataGoogle, 2017 · Active, term to 2037 · Inventors: Ion Constantinescu

  Claim 1 teaches a computing system associated with an existing knowledge graph receiving a third-party request to onboard entity data, matching the new entity to existing graph entities. This is the operational mechanism by which Google locks a domain to a Central Entity in its Knowledge Graph.

• US 9665643 B2 Knowledge-based entity detection and disambiguationMicrosoft, 2011 · Active, term to 2031 · Inventors: Kang Li, Yi Li, Yiping Zhou, Zhengdong Lv (+1 more)

  Teaches tagging named entities on web pages and using the disambiguated entity to organize search results (detailed description paragraphs on entity-based organization). Shows how a Central Entity signal cross-page is used by the ranking stack.

Google Content Warehouse leak 6 attributes

• RepositoryWebrefEntityAnnotations.topicalityScore "indicates how related is the entity to the overall topic of the document." The site-wide central entity is the one with highest aggregate topicality across pages.
• RepositoryWebrefEntityAnnotations.topicalityRank "Rank of the entity when sorted by topicality score." The top-ranked topicality entity on a site is the Central Entity in the leak's own terms.
• RepositoryWebrefSalientEntitySet (referenced by IndexingDocjoinerDataVersion.repositoryWebrefSalientEntitySet), the per-doc salient entity set, aggregated sitewide becomes the Central Entity set.
• RepositoryWebrefDetailedEntityScores.normalizedTopicality "the 'proportion' of the document that talks about the entity." Sitewide average is a quantitative proxy for Central Entity dominance.
• RepositoryWebrefEntityAnnotations.isImplicit flags whether the Central Entity is referenced implicitly (via query, URL, anchors) vs explicitly in body.
• GeostoreKnowledgeGraphReferenceProto.id "KG Identifier (MID)." The MID is the entity ID Koray treats as the "truth anchor" for Central Entity.

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Patent evidence 1 patent

• US 8423568 B2 Query classification using implicit labelsMicrosoft, 2009 · Active, term to 2029 · Inventors: Ariel D. Fuxman, Anitha Kannan, Andrew Brian Goldberg, Rakesh Agrawal

  Detailed-description paragraphs on commercial-intent category reinforcement. Shows that a user clicking the "Electronics" category reinforces commercial intent, the search-side mirror of a site asserting its own Source Context.

Google Content Warehouse leak 7 attributes

• CompressedQualitySignals.siteAuthority "converted from quality_nsr.SiteAuthority, applied in Qstar." Site-level authority is the leak's operational analog to Source Context trust.
• PerDocData.topPetacatTaxId "Top petacat of the site. Used in SiteboostTwiddler to determine result/query matching." Petacat is the site-level category classifier, exactly how Google decides what the site is about at the Source-Context level.
• IndexingDocjoinerDataVersion.qualityTwiddlerDomainClassification domain-level classification used by the ranking Twiddler system.
• QualityNsrNsrData.chardEncoded "Site-level chard score: site quality predictor based on content." Site-level content classification feeding Source Context.
• QualityNsrNsrData.tofu "Site-level tofu score: site quality predictor based on content."
• LocalWWWInfo.brickAndMortarStrength for local-business Source Contexts, measures how established the physical operation is.
• GeostoreEstablishmentProto.bizbuilderReference "indicates that a feature is claimed in CBDB (with the canonical state in MapFacts)." Business-claim status is a Source Context signal for local firms.

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Patent evidence 2 patents

• US 9870356 B2 Techniques for inferring the unknown intents of linguistic itemsMicrosoft, 2014 · Active, term to 2034 · Inventors: Dilek Hakkani-Tür, Fethiye Asli Celikyilmaz, Larry P. Heck, Gokhan Tur (+1 more)

  Abstract and claim 1: deterministically assigning intents to linguistic items (queries). This is the mechanism Google uses to collapse diverse surface queries ("get visa," "apply visa," "obtain visa") into a shared intent class, matching Koray's Central Search Intent predicate collapsing.

• US 11361227 B2 Onboarding of entity dataGoogle, 2017 · Active, term to 2037 · Inventors: Ion Constantinescu

  Claim 7 covers "attribute-to-entity associations," meaning the search engine binds verbs to a Central Entity, the operational substrate for [Predicate] + [Central Entity].

Google Content Warehouse leak 6 attributes

• KnowledgeAnswersIntentQueryLocalResultSignals.salientTermSet "Salient terms associated with this Local result based on the result gcids." The leak's model of intent is salient-terms-as-vectors.
• QualityPreviewSnippetDocumentFeatures.salientTermsScore salient terms score feeding snippet selection (snippet picks are intent-biased).
• KnowledgeAnswersIntentQueryLocalSignals.minSalientTermSimilarity "Minimum salient term similarity between Local results." Confirms Google clusters queries by salient-term similarity, which is the technical basis for Koray's "Central Search Intent" clustering.
• IndexingDocjoinerDataVersion.qualitySalientTermsSalientTermSet the per-doc salient term set used in retrieval.
• IndexingDocjoinerDataVersion.qualitySalientTermsQuerySmearingModel query smearing (expanding queries to adjacent variants), the mechanism behind unified search intent.
• RepositoryWebrefEntityAnnotations.explainedRangeInfo "All ranges explained by the entity or any other entity it implies. Used in the context of partial query interpretation." Direct support for intent-level entity resolution.

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Patent evidence 2 patents

• US 8868402 B2 Construction of text classifiersGoogle, 2009 · Active, term to 2029 · Inventors: Dmitry Korolev, Hartmut Maennel, Matthias Heiler, Michael Schaer (+3 more)

  Detailed description explicitly discusses collecting n+k-grams across a document set and using "Web site scores" to generate candidate phrases for a topic classifier. Figure 2 shows the n-gram extraction pipeline. Direct patent-side proof that Google uses site-level n-grams for topical classification, the engineering ancestor of siteFocus.

• US 7177796 B1 Automated set up of web-based natural language interfaceMicrosoft, 2000 · Expired 2020 · Inventors: Frederick J. Damerau, David E. Johnson

  Claim 1 and alternative claim: automatically inducing a topic hierarchy by examining a web site's structure and creating n-grams from pages associated with each topic. Grounds the idea that site structure plus n-gram distribution = site topic.

Google Content Warehouse leak 5 attributes

• QualityAuthorityTopicEmbeddingsVersionedItem.siteEmbedding "Compressed site/page embeddings." Site embedding is the mathematical compression of sitewide n-gram patterns.
• QualityAuthorityTopicEmbeddingsVersionedItem.siteFocusScore explicitly cited as siteFocus in Koray's writings; measures how tightly sitewide n-grams converge.
• QualityNsrNsrData.site2vecEmbeddingEncoded "Encoded site2vec embedding (to be used in superroot)." Direct site-level embedding storage.
• IndexingDocjoinerDataVersion.qualityNsrNsrSitechunksAnnotation and qualityNsrUvSitechunksAnnotation, sitechunk-level annotations where sitewide n-grams get persisted.
• PerDocData.nsrSitechunk "SiteChunk computed for nsr … In rare, but important cases it's based on page markup." The sitechunk is the unit sitewide n-grams aggregate on.

Related research and notes

• Content Warehouse API leak (May 2024): siteFocusScore and siteRadius properties documented in the leaked CompressedQualitySignals module. Not a patent but confirms the engineering concept Koray describes.

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Patent evidence 5 patents

• US 9158849 B2 Synthesis of webpage snippets using sub-pages of the webpageGoogle, 2013 · Active, term to 2033 · Inventors: Junli Xian, Harvey Jones, Samarth Keshava

  Claim 1 identifies sub-pages corresponding to a particular webpage within the same website, generates titles, and selects a subset for display. The mechanism that elevates a site's commercial core as its primary SERP representation.

• US 10776435 B2 Canonicalized online document sitelink generationGoogle, 2017 · Active, term to 2037 · Inventors: Vaibhav Vaish, Venky Ramachandran, David Philip Sisson, Ramakrishnan Kandhan (+3 more)

  Claim 1 teaches generating sitelinks from canonicalized internal structure. The core commercial folder (/germany/visa/) is literally what feeds Google's sitelinks surface, direct validation of concentrating commercial pages under one folder.

• US 7296016 B1 Point-of-view searchingGoogle, 2003 · Expired 2023 · Inventors: Martin Farach-Colton, Monika H. Henzinger, Bay-Wei Chang

  Teaches that rank flows from designated example URLs to all other documents easily reached from them. Applied internally, this is why the Core Section receives concentrated PageRank from the homepage and category hubs.

• US 8719276 B1 Ranking nodes in a linked database based on node independenceGoogle, 2011 · Active, term to 2031 · Inventors: Paul Haahr, Martin Kaszkiel, Amit Singhal

  Independent node contribution. Grounds why the Core Section benefits from diverse non-sitewide internal links rather than footer-link-farm patterns.

• US 9165040 B1 Producing a ranking for pages using distances in a web-link graphGoogle, 2006 · Expired 2026 · Inventors: Nissan Hajaj

  Claim 1 ranks pages by shortest distance from seed pages. The Core Section's shallow depth (/germany/visa/) is directly rewarded by this click-depth mechanic.

Google Content Warehouse leak 4 attributes

• PerDocData.homepagePagerankNs "The page-rank of the homepage of the site. Copied from the cdoc.doc().pagerank_ns() of the homepage." The Core Section typically inherits homepage PageRank through sculpting.
• AnchorsAnchorSource.homePageInfo "NOT_HOMEPAGE, NOT_TRUSTED, PARTIALLY_TRUSTED, and FULLY_TRUSTED." Classifies pages hub-authority; Core Sections sit near FULLY_TRUSTED hubs.
• IndexingDocjoinerAnchorStatistics.topPrOnsiteAnchorCount "anchor with pagrank > 51000 is the best anchor." Counts of highest-PageRank onsite anchors reaching the Core Section.
• GDocumentBase.PagerankNS "Pagerank-NearestSeeds is a pagerank score for the doc, calculated using NearestSeeds method. This is the production PageRank value teams should use."

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Patent evidence 3 patents

• US 7346839 B2 Information retrieval based on historical dataGoogle, 2003 · Expired 2023 · Inventors: Anurag Acharya, Matt Cutts, Jeffrey Dean, Paul Haahr (+6 more)

  Claim 1 enumerates document age, query-based click-through, and user-selection history as scoring inputs. The Outer Section's purpose is precisely to feed these signals.

• US 9158849 B2 Synthesis of webpage snippets using sub-pagesGoogle, 2013 · Active, term to 2033 · Inventors: Junli Xian, Harvey Jones, Samarth Keshava

  Internal link flow from Outer informational pages to Core commercial pages is the mechanism this patent references for sub-page scoring.

• US 8959093 B1 Ranking search results based on anchorsGoogle, 2010 · Active, term to 2030 · Inventors: Ajay P. Nerurkar

  Claim 1 propagates quality-of-result statistics (click behavior) from one document to another via a link whose anchor text is relevant to the query. Direct evidence that Outer Section user-click signals propagate upward to Core pages when internal anchors are query-relevant.

Google Content Warehouse leak 5 attributes

• QualityNavboostCrapsCrapsClickSignals.goodClicks + .badClicks, Outer Section pages exist to accumulate historical clicks; these are the exact signals they feed.
• QualityNavboostCrapsCrapsClickSignals.impressions Outer Section's role is to generate impressions to the Core Section.
• CompositeDocIndexingInfo.rawNavboost the per-URL raw click count supporting internal link transfer upward.
• AnchorsAnchor.isLocal "roughly indicates whether an anchor's source and target pages are on the same domain." The internal link bridge from Outer to Core sections.
• IndexingDocjoinerAnchorStatistics.topPrOnsiteAnchorCount direct measure of onsite link-equity flow Koray describes as "linking upward to the Core."

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Patent evidence 2 patents

• US 11886533 B2 Transferable neural architecture for structured data extraction from web documentsGoogle, 2020 · Active, term to 2040 · Inventors: Ying Sheng, Yuchen Lin, Sandeep Tata, Nguyen Vo

  Claims 1 and 9 extract structured data (node labels and node-pair relations) from DOM trees across websites via a transferable neural model. The engine that consumes an SCN into Google's structured record.

• US 9158849 B2 Synthesis of webpage snippets using sub-pagesGoogle, 2013 · Active, term to 2033 · Inventors: Junli Xian, Harvey Jones, Samarth Keshava

  The SERP surfacing payoff of a coherent SCN.

Google Content Warehouse leak 5 attributes

• RepositoryWebrefEntityAnnotations.segmentMentions "All mentions of a given concept grouped by segments." SCN is the engineering counterpart that maximizes dense entity segment-mentions sitewide.
• IndexingDocjoinerDataVersion.repositoryWebrefSalientEntitySet salient entity set per doc; SCN engineers content to maximize overlap of this set across inter-linked pages.
• IndexingDocjoinerDataVersion.indexingAnnotationsWaPassagesQueryToPassageAnnotations the query-to-passage mapping that an SCN engineers to be dense and complete.
• IndexingDocjoinerDataVersion.repositoryProtiumGoldmineTriplificationResult Goldmine's output of triple extraction from documents; SCN is explicit optimization for machine-readable triples.
• IndexingDocjoinerDataVersion.repositoryProtiumGoldmineTypeAssertionOutput type assertions extracted by Goldmine; SCN establishes consistent entity-type assertions sitewide.

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Google Content Warehouse leak 3 attributes

• CompressedQualitySignals.siteAuthority the single site-level authority signal into which all holistic signals are compressed.
• QualityNsrNsrData (40+ fields), site-level NSR aggregation across technical quality (chardEncoded), topic focus (localityScore, videoScore), click behavior (pnav, pnavClicks), freshness (nsrEpoch), content integrity (racterScores), link signals (siteLinkOut), browser signals (chromeInTotal).
• IndexingDocjoinerDataVersion.qualityQ2SiteQualityFeatures the Q2 site-quality feature aggregation module.

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Patent evidence 1 patent

• US 11361227 B2 Onboarding of entity dataGoogle, 2017 · Active, term to 2037 · Inventors: Ion Constantinescu

  Claim 7 attribute-to-entity binding: the Representative Source becomes the default attribute carrier for the Central Entity, structurally shadowing competitors.

Google Content Warehouse leak 6 attributes

• CompressedQualitySignals.authorityPromotion "authority promotion: converted from QualityBoost.authority.boost." The explicit boost applied to Representative Sources.
• CompressedQualitySignals.unauthoritativeScore "Unauthoritative score." Weaker sources carry an unauthoritative penalty; Representative Sources push them down.
• IndexingDocjoinerDataVersion.spamScamUnauthoritativeSite explicit flag for unauthoritative sites that lose rankings to Representative Sources.
• PerDocData.nsrIsCovidLocalAuthority and .nsrIsElectionAuthority, explicit "authority bit" whitelist mechanics; same pattern of Representative Source selection.
• IndexingDocjoinerDataVersion.localsearchAuthoritySiteAnnotation explicit authority-site annotation in local search.
• QualityCopiaFireflySiteSignal.latestBylineDateSec Firefly site signal captures Representative-Source freshness.

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Patent evidence 3 patents

• US 12393768 B2 Layout-aware multimodal pretrainingGoogle, 2020 · Active, term to 2040 · Inventors: Mingyang Zhang, Cheng Li, Tao Chen, Spurthi Amba Hombaiah (+3 more)

  Granted version of the above with the same multimodal document representation claims.

• US 7913163 B1 Determining semantically distinct regions of a documentGoogle, 2004 · Expired 2024 · Inventors: Yonatan Zunger

  Figure 2 shows a chunk-tree hierarchical data structure of a web page. Grounds segmentation of a page into visual-semantic chunks (pixels) that are scored independently from the text (letters).

• US 9069855 B2 Modifying a hierarchical data structure according to a pseudo-rendering of a structured documentGoogle, 2011 · Active, term to 2031 · Inventors: Yonatan Zunger

  Teaches "pseudo-rendering" during indexing, Google computes rendered layout before scoring. The pixel-geometry side of the framework is explicitly modeled.

Google Content Warehouse leak 9 attributes

• Letters (text)
• QualityPreviewSnippetDocumentFeatures.salientTermsScore .salientPositionBoostScore, .schemaOrgDescriptionBoostScore, the "letters" scoring layer at snippet-selection time.
• DocProperties.numTokens "The number of tokens, tags and punctuations in the tokenized contents."
• Bytes (rendering/server)
• CompositeDocAdditionalChecksums.NoTransientChecksum96 "Same as ContentChecksum96 but without transient boilerplate." Direct boilerplate accounting that affects byte cost.
• IndexingDocjoinerDataVersion.boilerplateVisibleBytesEstimates "boilerplate visible bytes estimates." Google literally measures boilerplate byte footprint.
• ContentAwareCropsIndexing.mustangBytes "Compact representation for Mustang storage." Mustang serving stores a byte-compact doc rep.
• Pixels (layout/visual)
• IndexingDocjoinerDataVersion.indexingMlPageStructure ML-based page-structure annotation.
• IndexingDocjoinerDataVersion.fontsAnalysisAnnotatorFontsAnnotation font analysis annotation, confirming typography-level ("pixels") parsing.
• IndexingDocjoinerDataVersion.repositoryAnnotationsRdfaBreadcrumbs and .richsnippetsPageMap, layout-structure annotations.
• HtmlrenderWebkitHeadlessProtoImage module, headless render metadata, the "pixels" input.

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Patent evidence 1 patent

• US 11886533 B2 Transferable neural architecture for structured data extraction from web documentsGoogle, 2020 · Active, term to 2040 · Inventors: Ying Sheng, Yuchen Lin, Sandeep Tata, Nguyen Vo

  Claim 1 extracts structured data by processing DOM nodes with content and XPath, fully formalizing the cheap-extraction premise. Pages with clean DOM and short XPaths are cheap to extract; noisy pages are expensive and bypassed.

Google Content Warehouse leak 5 attributes

• CompositeDocQualitySignals.lastSignificantUpdate recency gate (ChatGPT-style recency bias).
• QualityNsrPQData.contentEffort "LLM-based effort estimation for article pages." Direct effort scoring that discards low-effort pages at the recruitment gate.
• QualityPreviewSnippetDocumentFeatures.leadingtextDistanceScore distance of answer from top of document. LLM agents reward answers near the top.
• QualityPreviewSnippetExperimentalFeatures.isLikelyHomepage explicit homepage detection; LLM agents may deprioritize homepage boilerplate.
• QualityPreviewSnippetBrainFeatures.modelScore SnippetBrain model score estimating passage usefulness.

Related research and notes

• REALM: Retrieval-Augmented Language Model Pre-Training (Guu et al., ICML 2020, arXiv:2002.08909). Grounds the "retrieval before generation" premise and the top-k cutoff that forces LLMs to apply strict cost gates before passage synthesis.
• FreshLLMs (Vu et al., 2023, arXiv:2310.03214). Documents how recency filters materially change which pages are selected for RAG passages, matching Koray's "ChatGPT applies aggressive recency checks" observation.

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Google Content Warehouse leak 4 attributes

• IndexingDocjoinerDataVersion.repositoryProtiumGoldmineTriplificationResult Goldmine's triple extractor literally produces [Entity + Attribute + Value] triples from documents. The Raw Topical Map is the human-authored input for this extractor.
• GeostoreAttributeProto.enumIdValue attribute values stored as explicit enum IDs on entities, the production store for E-A-V-style raw topic nodes in Geo.
• GeostoreAttributeIdProto.id attribute identifier; direct evidence of attribute-ID storage schemes.
• KnowledgeGraphDateTimeProto explicit storage for date-attribute values on KG entities.

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Patent evidence 3 patents

• US 9158849 B2 Synthesis of webpage snippets using sub-pages of the webpageGoogle, 2013 · Active, term to 2033 · Inventors: Junli Xian, Harvey Jones, Samarth Keshava

  Claim 1: identify sub-pages, determine titles, and select a subset for display. Processed title tags per sub-page is the output artifact Google itself reads to understand a processed topical map. Reference: Claim 1 steps (a)-(c).

• US 10776435 B2 Canonicalized online document sitelink generationGoogle, 2017 · Active, term to 2037 · Inventors: Vaibhav Vaish, Venky Ramachandran, David Philip Sisson, Ramakrishnan Kandhan (+3 more)

  Claim 1 teaches sitelink generation from a canonicalized online document structure. The processed map's URL slug conventions (/germany/visa/passport/) feed directly into this sitelink mechanism. Reference: Claim 1.

• US 11886533 B2 Transferable neural architecture for structured data extraction from web documentsGoogle, 2020 · Active, term to 2040 · Inventors: Ying Sheng, Yuchen Lin, Sandeep Tata, Nguyen Vo

  Claims 1 and 9 extract structured data from fully executed DOM trees (Title Tags, headings, content) across websites. Processed topical maps feed cleaner input into this extractor than raw text. Reference: Claim 1 (steps A-F), Claim 9.

Google Content Warehouse leak 4 attributes

• QualityPreviewRanklabTitle.goldmineSalientTermFactor title-level salient-term factor used in SERP title selection. The Processed Topical Map's title is scored against this.
• QualityPreviewSnippetDocumentFeatures.experimentalTitleSalientTermsScore explicit title-vs-salient-terms score.
• IndexingDocjoinerDataVersion.indexingUrlPatternUrlTreeUrlPatternAnnotations URL-tree pattern annotations; the "processed URL slug" layer of the Processed Topical Map.
• IndexingSignalAggregatorUrlPatternSignals.pagerankScore and .coverage, URL-pattern-level PageRank and coverage, the directional output of a well-processed map.

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Patent evidence 5 patents

• US 9165040 B1 Producing a ranking for pages using distances in a web-link graphGoogle, 2006 · Expired 2026 · Inventors: Nissan Hajaj

  Claim 1 ranks pages by shortest distance from seed pages. The Information Tree's Root → Seed → Node hierarchy is directly rewarded by this click-depth mechanic because each level stays shallow. Reference: Claim 1, Fig. 2.

• US 9953049 B1 Producing a ranking for pages using distances in a web-link graphGoogle, 2015 · Active, term to 2035 · Inventors: Nissan Hajaj

  Continuation extending the seed-distance ranking at scale. Reference: Claim 1.

• US 8825646 B1 Scalable system for determining short paths within web link networkGoogle, 2009 · Active, term to 2029 · Inventors: Jesse Louis Alpert, Nissan Hajaj

  Shards the directed graph for parallel shortest-path computation. The infrastructure that evaluates an Information Tree at web scale. Reference: Abstract.

• US 10776435 B2 Canonicalized online document sitelink generationGoogle, 2017 · Active, term to 2037 · Inventors: Vaibhav Vaish, Venky Ramachandran, David Philip Sisson, Ramakrishnan Kandhan (+3 more)

  Clean canonical hierarchy is a direct prerequisite for sitelinks. Reference: Claim 1.

• US 9158849 B2 Synthesis of webpage snippets using sub-pagesGoogle, 2013 · Active, term to 2033 · Inventors: Junli Xian, Harvey Jones, Samarth Keshava

  Sub-page identification depends on the Information Tree's clarity. Reference: Claim 1 step (a).

Google Content Warehouse leak 5 attributes

• IndexingDocjoinerDataVersion.qualityBreadcrumbsBreadcrumbAnnotations breadcrumb annotation pipeline. Breadcrumbs are the explicit Information Tree annotation layer.
• RepositoryAnnotationsRdfaBreadcrumbs.crumb and .url, "Each crumb represents one link of the breadcrumb chain." Direct storage of Information Tree paths.
• RepositoryAnnotationsRdfaCrumb.title and .url, the crumb-level title and URL at each tree level.
• IndexingDocjoinerDataVersion.indexingUrlPatternUrlTreeUrlPatternAnnotations explicit "URL tree" nomenclature in the leak.
• IndexingMobileInterstitialsProtoDesktopInterstitials.urlTree "URL tree of interstitial patterns … used as site-level signal."

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Patent evidence 2 patents

• US 11886533 B2 Transferable neural architecture for structured data extraction from web documentsGoogle, 2020 · Active, term to 2040 · Inventors: Ying Sheng, Yuchen Lin, Sandeep Tata, Nguyen Vo

  Claims 1 and 9: extracts machine-actionable structured data (node labels and pair relations) from HTML. The E-A-V triple is the output format of this patent's extraction pipeline. Reference: Claim 1 step E (node-label generation), Claim 9 (pair-label generation).

• US 11361227 B2 Onboarding of entity dataGoogle, 2017 · Active, term to 2037 · Inventors: Ion Constantinescu

  Claim 1: computing system associated with an existing knowledge graph receives entity-attribute data for onboarding. The direct Google mechanism that ingests site-asserted E-A-V triples. Reference: Claim 1.

Google Content Warehouse leak 5 attributes

• IndexingDocjoinerDataVersion.repositoryProtiumGoldmineTriplificationResult the actual triple extractor. E-A-V is the input shape.
• GeostoreAttributeIdProto.id + GeostoreAttributeProto.enumIdValue, Google's own E-A-V storage for Geo entities.
• RepositoryWebrefEntityAnnotations.detailedEntityScores "Additional information about how the entity relates to the page." Entity-level attribute scoring per document.
• IndexingDocjoinerDataVersion.repositoryProtiumGoldmineTypeAssertionOutput type assertions (Entity is-a Type) extracted from text.
• GoogleCloudContentwarehouseV1Property.propertyValues and GoogleCloudContentwarehouseV1PropertyDefinition.propertyTypeOptions, explicit property-value structured storage on the Cloud Content Warehouse API itself.

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Patent evidence 1 patent

• US 11361227 B2 Onboarding of entity dataGoogle, 2017 · Active, term to 2037 · Inventors: Ion Constantinescu

  Claim 7 (attribute-to-entity associations) implies attribute prioritization is a required step before onboarding into the knowledge graph: Google will not accept noisy, undifferentiated attribute dumps. Reference: Claim 7.

Google Content Warehouse leak 3 attributes

• RepositoryWebrefDetailedEntityScores.normalizedTopicality "the 'proportion' of the document that talks about the entity." Attributes expressed through high-topicality entities get Root-tier weight.
• RepositoryWebrefEntityAnnotations.topicalityRank rank-ordering of entities per document; prioritization lives here.
• RepositoryWebrefAggregatedEntityNameScores.entityScore aggregated per-entity-name score across the web; "Unique" attributes benefit from distinctive high-score names.

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Patent evidence 1 patent

• US 7853589 B2 Web spam page classification using query-dependent dataMicrosoft, 2007 · Active, term to 2027 · Inventors: Krysta Svore, Chris Burges

  Granted version of US-2008270376-A1.

Google Content Warehouse leak 4 attributes

• RepositoryWebrefEntityAnnotations.topicalityScore relevance to document topic. Filters low-relevance attributes.
• RepositoryWebrefEntityAnnotations.confidenceScore "the overall confidence that the entity is annotated." Low-confidence attributes get filtered.
• RepositoryWebrefAnnotationStatsPerType.numRangesWithCandidates count of ranges with annotation candidates per segment type; filtration criterion.
• RepositoryWebrefAnnotationStatsPerType.avgOpenWorld "The average score for the open world." Attributes below average get filtered.

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Google Content Warehouse leak 3 attributes

• IndexingDocjoinerDataVersion.indexingAnnotationsWaPassagesQueryToPassageAnnotations the actual query-to-passage mapping Google maintains. 3-Column analysis engineers content for this mapping.
• IndexingDocjoinerDataVersion.indexingAnnotationsWaPassagesTrimmedQueryToPassageAnnotations trimmed query-to-passage, the compressed version Google ships to serving.
• QualityPreviewSnippetDocumentFeatures.salientTermsScore target queries ultimately resolve to salient-terms. Column 2 (benchmark terminology) is the gold vocabulary for this score.

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Patent evidence 5 patents

• US 7346839 B2 Information retrieval based on historical dataGoogle, 2003 · Expired 2023 · Inventors: Anurag Acharya, Matt Cutts, Jeffrey Dean, Paul Haahr (+6 more)

  Enumerates document age, link growth rate, and content update rate as scoring factors. Momentum = all these signals in the positive direction. Reference: Claim 1 and claims 10-38.

• US 9361369 B1 Clustering news online content based on content freshness and quality of content sourceGoogle, 2012 · Active, term to 2032 · Inventors: Krishna Bharat, Michael Curtiss, Michael Schmitt

  Claim 1 scores online content based on freshness. Reference: Claim 1.

• US 8407204 B2 Minimizing visibility of stale content in web searching including revising web crawl intervalsGoogle, 2011 · Active, term to 2031 · Inventors: Anton P. T. Carver

  Detailed description formula: P_stale, PR_pagerank, T_click_rate, T_content_update_rate combined to compute crawl interval. The content update rate is the direct Momentum signal; higher update rate shortens the crawl interval. Reference: Detailed description formula.

• US 8549014 B2 Document scoring based on document content updateGoogle, 2011 · Active, term to 2031 · Inventors: Anurag Acharya, Jeffrey Dean, Paul Haahr, Monika Henzinger (+3 more)

  Scores documents based on anchor-text and content update rate. Momentum signal confirmed. Reference: Claim 1.

• US 8407231 B2 Document scoring based on link-based criteriaGoogle, 2010 · Active, term to 2030 · Inventors: Anurag Acharya, Matt Cutts, Jeffrey Dean, Paul Haahr (+4 more)

  Timed appearance of each inbound link; a rising-link cadence rewards Momentum. Reference: Claim 1.

Google Content Warehouse leak 5 attributes

• QualityAuthorityTopicEmbeddingsVersionedItem.siteRadius Vastness measured as site_embedding spread.
• RepositoryWebrefAnnotatorProfile.numEntities .numMentions, .numTokens, the raw counts that quantify Depth (more entities + mentions per doc).
• IndexingDocjoinerDataVersion.firstseen first-seen date for crawl; Momentum = consistent new firstseen dates.
• AnchorsAnchor.firstseenDate and .firstseenNearCreation, link-level freshness used to measure publishing momentum via inlink velocity.
• CompositeDocQualitySignals.lastSignificantUpdate content-update recency, the content-side Momentum signal.

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Patent evidence 4 patents

• US 9158849 B2 Synthesis of webpage snippets using sub-pages of the webpageGoogle, 2013 · Active, term to 2033 · Inventors: Junli Xian, Harvey Jones, Samarth Keshava

  Claim 1 identifies sub-pages and generates titles, the engine-side mechanism that rewards treating a query as a sub-page (QDP). Reference: Claim 1.

• US 10776435 B2 Canonicalized online document sitelink generationGoogle, 2017 · Active, term to 2037 · Inventors: Vaibhav Vaish, Venky Ramachandran, David Philip Sisson, Ramakrishnan Kandhan (+3 more)

  Sitelink generation keys off the canonical sub-page URL. Queries that deserve a page become sitelink targets. Reference: Claim 1.

• US 7913163 B1 Determining semantically distinct regions of a documentGoogle, 2004 · Expired 2024 · Inventors: Yonatan Zunger

  Chunk-tree segmentation of a page into semantic blocks, the direct engineering mechanism that reads a heading section (QDH) as a distinct unit from the surrounding page. Reference: Fig. 2 (chunk tree).

• US 9069855 B2 Modifying a hierarchical data structure according to a pseudo-rendering of a structured documentGoogle, 2011 · Active, term to 2031 · Inventors: Yonatan Zunger

  Pseudo-rendering annotates nodes according to visual hierarchy, allowing an H2 (QDH) to be scored as a sub-document. Reference: Fig. 2.

Google Content Warehouse leak 3 attributes

• IndexingDocjoinerDataVersion.indexingAnnotationsWaPassagesQueryToPassageAnnotations query-to-passage; a QDH heading becomes a passage anchor here.
• QualityPreviewSnippetDocumentFeatures.leadingtextDistanceScore a QDS single-sentence answer scores best when near the top (low leading-text distance).
• IndexingSignalAggregatorAgeWeightedCoverageData age-weighted coverage data; QDP pages accumulate standalone coverage.

Related research and notes

• Google patent US-9213748-B1, Methods for selecting and presenting content based on an understanding of user interests (Google, 2013, PAA family). Ranks sub-questions beneath a primary query, the PAA surface rewards QDH-styled sub-section coverage.

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Patent evidence 1 patent

• US 11361227 B2 Onboarding of entity dataGoogle, 2017 · Active, term to 2037 · Inventors: Ion Constantinescu

  Claim 1's attribute-matching logic implies tolerance thresholds that themselves shift as the graph evolves, a border mechanism. Reference: Claim 1.

Google Content Warehouse leak 3 attributes

• QualityAuthorityTopicEmbeddingsVersionedItem.versionId versioning on topic embeddings is Google's explicit acknowledgment that topic boundaries shift between versions.
• CompressedQualitySignals.topicEmbeddingsVersionedData "Versioned TopicEmbeddings data to be populated later into superroot / used directly in scorers." Multiple versions coexist, confirming re-calibration.
• QualityNsrNsrData.nsrEpoch "The epoch from which this NSR value is coming from." Explicit epoch versioning; borders shift between epochs.

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Google Content Warehouse leak 3 attributes

• Prominence → RepositoryWebrefDetailedEntityScores.normalizedTopicality (how prominent the entity is on the page)
• Relevance → RepositoryWebrefEntityAnnotations.topicalityScore (overall-topic relevance)
• Popularity → RepositoryWebrefAggregatedEntityNameScores.entityScore (aggregated name-level popularity across the index)

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Google Content Warehouse leak 2 attributes

• RepositoryWebrefEntityAnnotations.isResolution "True if the entity is an MDVC summary entity, i.e. it might not be mentioned directly on the query, but it is the product of resolving a set of explicit annotations." This is literally compound-attribute resolution.
• RepositoryWebrefEntityAnnotations.explainedRangeInfo ranges explained by an entity "or any other entity it implies." Compound attribute inference is explicit here.

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Patent evidence 7 patents

• US 9158849 B2 Synthesis of webpage snippets using sub-pages of the webpageGoogle, 2013 · Active, term to 2033 · Inventors: Junli Xian, Harvey Jones, Samarth Keshava

  Claim 1: identifies sub-pages corresponding to a particular webpage within the same website. Direct patent on recognizing sub-site hubs within a parent site. Reference: Claim 1 steps (a)-(c).

• US 10776435 B2 Canonicalized online document sitelink generationGoogle, 2017 · Active, term to 2037 · Inventors: Vaibhav Vaish, Venky Ramachandran, David Philip Sisson, Ramakrishnan Kandhan (+3 more)

  Sitelink generation requires identifying coherent sub-site hubs under one brand. Reference: Claim 1.

• US 8423551 B1 Clustering internet resourcesGoogle, 2010 · Active, term to 2030 · Inventors: Aner Ben-Artzi, Kirill Buryak, Glenn M. Lewis, Jun Peng (+1 more)

  Claim 1: receives a user request referencing keywords, identifies content items, and clusters them. Grounds the ability to collect sub-site content into a themed cluster for retrieval. Reference: Claim 1.

• US 9483524 B1 Ranking communities based on contentGoogle, 2013 · Active, term to 2033 · Inventors: Bryan Christopher Horling, Okan Kolak

  Ranks sub-communities by content, a social-network parallel to subfolder evaluation. Grounds the mechanism of treating a subset of URLs as a distinct ranked collection. Reference: Abstract.

• US 8886655 B1 Visual display of topics and content in a map-like interfaceGoogle, 2012 · Active, term to 2032 · Inventors: Palash Nandy, Michael Sho Ogawa

  Claim 1 identifies topics from metadata and determines relationships based on past user interaction, the sub-topic retrieval primitive. Reference: Claim 1.

• US 8868402 B2 Construction of text classifiersGoogle, 2009 · Active, term to 2029 · Inventors: Dmitry Korolev, Hartmut Maennel, Matthias Heiler, Michael Schaer (+3 more)

  "Web site scores" used to generate candidate phrases, the site-level analog that applies to a subfolder's distinct scoring. Reference: Detailed description.

• US 9165040 B1 Producing a ranking for pages using distances in a web-link graphGoogle, 2006 · Expired 2026 · Inventors: Nissan Hajaj

  The shortest-path mechanic measures sub-site distance from site-level seeds, subfolder depth is a direct input. Reference: Claim 1.

Google Content Warehouse leak 3 attributes

• QualityNsrNsrData.siteChunk + .siteChunkSource, "Primary NSR sitechunk … In rare, but important cases it's based on page markup." Confirms subfolders can function as independent sitechunks for NSR.
• PerDocData.nsrSitechunk "SiteChunk computed for nsr. It some cases it can use more information than just url (e.g. youtube channels)." Direct multi-level sub-site chunking.
• QualityNsrNsrData.nsrdataFromFallbackPatternKey "If true indicates that we do not have NSR data computed for the chunk, and instead the data is coming from an average of other host chunks." Subfolders need their own NSR data to escape host average, which is topic distillation.

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Patent evidence 2 patents

• US 11361227 B2 Onboarding of entity dataGoogle, 2017 · Active, term to 2037 · Inventors: Ion Constantinescu

  Claim 1: computing system associated with an existing knowledge graph. The existing knowledge graph is itself the Golden Source, new submissions are checked against it. Reference: Claim 1.

• US 9665643 B2 Knowledge-based entity detection and disambiguationMicrosoft, 2011 · Active, term to 2031 · Inventors: Kang Li, Yi Li, Yiping Zhou, Zhengdong Lv (+1 more)

  Disambiguates entities against a trusted knowledge base; the trusted knowledge base is the Golden Source. Reference: Detailed description.

Google Content Warehouse leak 4 attributes

• IndexingDocjoinerDataVersion.repositoryAnnotationsSporeAuthoritySignal "Spore authority signal" from the structured-data pipeline that feeds Knowledge Graph. Spore outputs feed Google's own golden-source consensus.
• IndexingDocjoinerDataVersion.researchStructuredDataKnowledgeExplorationAnnotations explicit research-level structured-data annotations used for Golden-Source baselines.
• GeostoreTrustSignalsProto.sourceTrust "Trust signals for the source of a given observation, typically based on historical evidences or status (like internal Google operator)." Trust hierarchy for golden sources.
• GeostoreSourceTrustProto.level "The level of trust for the source of the observation."

Related research and notes

• Luna Dong, Knowledge Vault / Knowledge-Based Trust (KBT) (Dong et al., KDD 2014; also arXiv:1502.03519, "Knowledge-Based Trust: Estimating the Trustworthiness of Web Sources"). Introduces KBT: scoring web sources by factual agreement with a trusted knowledge graph. The academic foundation for Golden Sources.
• REALM (Guu et al., ICML 2020, arXiv:2002.08909), retrieval-augmented generation retrieves from a pretrained Golden-Source corpus before generating.
• Attributed Question Answering (Bohnet et al., arXiv:2212.08037), grounds answer attribution to Golden Sources.

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Patent evidence 3 patents

• US 11983173 B2 Systems and methods for generation and application of schema-agnostic query templatesGoogle, 2019 · Active, term to 2039 · Inventors: Anastasios Kementsietsidis, Jay Yogeshbhai Pandya, Tingting TANG, Laurren Kanner

  Direct hit. Abstract: "generate query templates that are expressed in a generic schema-agnostic language. The query templates can be generated 'from scratch' or can be automatically [applied across entities]." This is Google's own patent for the Query Template concept Koray describes. Reference: Abstract and claim 1.

• US 7647298 B2 Generation of query and update views for object relational mappingMicrosoft, 2007 · Active, term to 2027 · Inventors: Atul Adya, Philip A. Bernstein, Sergey Melnik

  Mapping-language approach supporting table-per-hierarchy and table-per-type patterns, the schema-level foundation of query templates. Reference: Detailed description on mapping language.

• US 11361227 B2 Onboarding of entity dataGoogle, 2017 · Active, term to 2037 · Inventors: Ion Constantinescu

  Attribute-to-entity binding across many entities is the structural form of a Query Template operating on an entity class. Reference: Claim 7.

Google Content Warehouse leak 4 attributes

• IndexingDocjoinerDataVersion.qualityNavboostCrapsCrapsPatternSignal "Craps pattern signal." A template-level click pattern aggregated across entity-attribute variations.
• CompositeDocIndexingInfo.urlPatternSignals "UrlPatternSignals for this doc, used to compute document score in LTG." URL patterns are the serving-side counterpart to query templates.
• IndexingDocjoinerDataVersion.indexingSignalAggregatorUrlPatternSignals URL-pattern signal aggregation; the pipeline that persists template authority.
• CountryCountryAttachment.urlPatternBasedCountry country attribution via URL pattern; template authority stays per-country.

Related research and notes

• Dong et al., Knowledge Vault (KDD 2014). Combines web-extracted triples with query-template-like patterns to populate the KG.

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Patent evidence 4 patents

• US 11960554 B2 Unsupervised entity and intent identification for improved search query relevanceMicrosoft, 2021 · Active, term to 2041 · Inventors: Gautam Prasad, Varun Appaswami, Bhanu Teja Chunduri

  Abstract: improves search query relevance by extracting entity + intent together from search-page data. Reads the multi-attribute compound ("which software for QR code generators: design, style, speed") as a joint entity-intent extraction. Reference: Abstract and claim 1.

• US 11068554 B2 Unsupervised entity and intent identificationMicrosoft, 2019 · Active, term to 2039 · Inventors: Gautam Prasad, Varun Appaswami, Bhanu Teja Chunduri

  Granted earlier version.

• US 11983173 B2 Schema-agnostic query templatesGoogle, 2019 · Active, term to 2039 · Inventors: Anastasios Kementsietsidis, Jay Yogeshbhai Pandya, Tingting TANG, Laurren Kanner

  Templates with multiple variable slots are literally the "Three-Top" prompt pattern expressed as a schema. Reference: Abstract.

• US 11361227 B2 Onboarding of entity dataGoogle, 2017 · Active, term to 2037 · Inventors: Ion Constantinescu

  Attribute-to-entity binding allows multi-attribute compound facts. Reference: Claim 7.

Google Content Warehouse leak 3 attributes

• IndexingDocjoinerDataVersion.indexingAnnotationsWaPassagesQueryToPassageAnnotations multi-attribute compound queries resolve against passage annotations. The 3-top approach engineers for this exact mapping.
• QualityPreviewRanklabTitle.goldmineSalientTermFactor multi-attribute compound queries evaluated against title salient-term factor.
• RepositoryWebrefEntityAnnotations.explainedRangeInfo compound-attribute ranges explained by implied entities, the matching layer for 3-top queries.

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Patent evidence 1 patent

• US 8886747 B1 Verifying domain ownershipGoogle, 2010 · Active, term to 2030 · Inventors: Bryan CHEN, Erik M. Kline, Marija Mikic-Rakic

  Claim 1: verifies that an entity controls content associated with a domain name. The ownership-verification primitive that underlies "Topical Ownership" as a measurable, enforced state. Reference: Claim 1.

Google Content Warehouse leak 4 attributes

• CompressedQualitySignals.authorityPromotion the explicit authority-boost assigned to "owned" sources.
• QualityNsrNsrData.siteAutopilotScore "Aggregated value of url autopilot scores for this sitechunk." Autopilot (promoted/trusted) status at site chunk level.
• IndexingDocjoinerDataVersion.localsearchAuthoritySiteAnnotation explicit "authority site" annotation.
• PerDocData.nsrIsCovidLocalAuthority .nsrIsElectionAuthority, authority whitelist bits (precedent for how Topical Ownership is codified for whitelisted verticals).

Related research and notes

• Dong et al., Knowledge Vault / KBT (KDD 2014), Topical Ownership operationalized as the knowledge-base trust score of a domain.

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Patent evidence 6 patents

• US 9448992 B2 Natural language search results for intent queriesGoogle, 2013 · Active, term to 2033 · Inventors: Tomer Shmiel, Dvir Keysar, Yonatan Erez

  Claim 1: parses documents from authoritative sources to generate heading-text pairs; associates each pair with an intent template. The H1-to-text pairing is literally how Google reads Macro Context off a page. Reference: Claim 1.

• US 9959315 B1 Context scoring adjustments for answer passagesGoogle, 2014 · Active, term to 2034 · Inventors: Nitin Gupta, Srinivasan Venkatachary, Lingkun Chu, Steven D. Baker

  Claim 1 and dependent claims construct a heading vector from H1 downward and score passages against it. The H1 anchors the Macro Context in this patent's heading vector. Reference: Claim 1, dependent claims on heading vectors.

• US 10354292 B1 Systems and methods for generating navigation filtersGoogle, 2016 · Active, term to 2036 · Inventors: Shreyas Doshi, Venky Ramachandran, Gaurav Garg, Poorva Hari Potdar (+3 more)

  Detailed description with explicit HTML example (<h2>Featured Brands</h2>, <h2>Special Offers</h2>) shows Google parses heading structure to derive navigation filters. Each heading cluster is a Macro Context block. Reference: Detailed description HTML example.

• US 7913163 B1 Determining semantically distinct regions of a documentGoogle, 2004 · Expired 2024 · Inventors: Yonatan Zunger

  Chunk-tree segmentation treats the primary (H1-anchored) chunk as the dominant semantic region. Reference: Fig. 2.

• US 9069855 B2 Modifying a hierarchical data structure according to a pseudo-renderingGoogle, 2011 · Active, term to 2031 · Inventors: Yonatan Zunger

  Pseudo-rendering elevates the top-of-document hierarchy, the top 70% weighting Koray describes. Reference: Fig. 2.

• US 9158849 B2 Synthesis of webpage snippets using sub-pagesGoogle, 2013 · Active, term to 2033 · Inventors: Junli Xian, Harvey Jones, Samarth Keshava

  Sub-page title selection assumes one dominant title (Macro Context) per URL. Reference: Claim 1.

Google Content Warehouse leak 3 attributes

• QualityPreviewSnippetDocumentFeatures.salientTermsScore the primary salient-terms score, driven by the single Macro Context.
• RepositoryWebrefEntityAnnotations.topicalityRank (rank 1 entity = the Macro Context's primary entity).
• IndexingDocjoinerDataVersion.repositoryWebrefSalientEntitySet if the salient-entity set has one dominant entity, Macro Context is clean; multiple competing entities = dilution.

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Patent evidence 4 patents

• US 9959315 B1 Context scoring adjustments for answer passagesGoogle, 2014 · Active, term to 2034 · Inventors: Nitin Gupta, Srinivasan Venkatachary, Lingkun Chu, Steven D. Baker

  Heading vector mechanic recognizes subordinate headings as context-shifted passages, the engineering form of Micro Context. Reference: Dependent claims on subordinate text.

• US 7913163 B1 Determining semantically distinct regions of a documentGoogle, 2004 · Expired 2024 · Inventors: Yonatan Zunger

  Chunk-tree separates subordinate regions from the dominant one. Reference: Fig. 2.

• US 9448992 B2 Natural language search resultsGoogle, 2013 · Active, term to 2033 · Inventors: Tomer Shmiel, Dvir Keysar, Yonatan Erez

  Secondary heading-text pairs are the mechanical form of Micro Context sections. Reference: Claim 1.

• US 8959093 B1 Ranking search results based on anchorsGoogle, 2010 · Active, term to 2030 · Inventors: Ajay P. Nerurkar

  Internal anchors from Micro Context sections to primary pages propagate click evidence conditioned on query relevance, the exact mechanism by which Micro Context creates natural internal linking justification. Reference: Claim 1.

Google Content Warehouse leak 3 attributes

• RepositoryWebrefEntityAnnotations.segmentMentions "All mentions of a given concept grouped by segments … Webref, there are many different kinds of segment, such as content, title and anchors." Micro contexts persist as secondary segments.
• RepositoryWebrefEntityAnnotations.topicalityRank ranks < 1 are the Micro Context entities on the page.
• AnchorsAnchor micro-context justified internal links are stored as anchors with their own context text.

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Patent evidence 4 patents

• US 9959315 B1 Context scoring adjustments for answer passagesGoogle, 2014 · Active, term to 2034 · Inventors: Nitin Gupta, Srinivasan Venkatachary, Lingkun Chu, Steven D. Baker

  Dependent claims explicitly describe "text section that is subordinate to a respective heading when the text section includes a text formatted according to [a non-heading style]." The patent names First Subordinate Text as a scorable unit. Reference: Dependent claims specifying subordinate text formatting.

• US 9448992 B2 Natural language search results for intent queriesGoogle, 2013 · Active, term to 2033 · Inventors: Tomer Shmiel, Dvir Keysar, Yonatan Erez

  Claim 1: "parses documents from authoritative sources to generate heading-text pairs." The "text" half is First Subordinate Text; the pair is extracted together into an intent template. Reference: Claim 1.

• US 9323827 B2 Identifying key terms related to similar passagesGoogle, 2008 · Active, term to 2028 · Inventors: William N. Schilit, Okan Kolak

  Identifies pre-context and post-context for a passage, First Subordinate Text sits in the pre-context window of the heading. Reference: Claim-method extracting pre-context / post-context.

• US 9946695 B2 Automatically generating content layout based on selected highest scored image and selected text snippetGoogle, 2013 · Active, term to 2033 · Inventors: Guannan Zhang, Kai Ye, Yongtai Zhu, Zuo Yan (+4 more)

  Selects a highest-scored text snippet from a landing resource, pages where the first subordinate text is a clean declarative sentence win. Reference: Claim 1.

Google Content Warehouse leak 3 attributes

• QualityPreviewSnippetDocumentFeatures.leadingtextDistanceScore "leading text distance" scoring. Direct measurement of how close the primary answer is to the start.
• QualityPreviewSnippetDocumentFeatures.salientPositionBoostScore "salient position boost score." Rewards salient terms near the top (First Subordinate Text).
• QualityPreviewSnippetBrainFeatures.modelScore SnippetBrain evaluates lead-of-document text to score passage quality.

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Patent evidence 2 patents

• US 9959315 B1 Context scoring adjustments for answer passagesGoogle, 2014 · Active, term to 2034 · Inventors: Nitin Gupta, Srinivasan Venkatachary, Lingkun Chu, Steven D. Baker

  Heading vectors preserve context across heading transitions, a Contextual Bridge heading maintains the vector between unrelated sub-topics. Reference: Dependent claims.

• US 8959093 B1 Ranking search results based on anchorsGoogle, 2010 · Active, term to 2030 · Inventors: Ajay P. Nerurkar

  Query-relevant anchors propagate quality signals, Contextual Bridges give the editorial justification for the internal anchor. Reference: Claim 1.

Google Content Warehouse leak 2 attributes

• RepositoryWebrefEntityAnnotations.segmentMentions bridging micro-sections carry entity mentions tied to internal link targets.
• AnchorsAnchor and IndexingDocjoinerAnchorStatistics, the internal link each bridge enables is persisted here; the quality of the bridge is measured via link quality buckets.

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Patent evidence 6 patents

• US 9959315 B1 Context scoring adjustments for answer passagesGoogle, 2014 · Active, term to 2034 · Inventors: Nitin Gupta, Srinivasan Venkatachary, Lingkun Chu, Steven D. Baker

  Claim 1 and dependent claims: "the heading vector includes text of each heading in the path in the heading hierarchy from the root heading to the respective heading." This is the exact Contextual Vector Koray describes, formalized as a mathematical vector of heading texts from H1 downward. Reference: Dependent claim on heading-vector match scores.

• US 11409748 B1 Context scoring adjustments for answer passagesGoogle, 2018 · Active, term to 2038 · Inventors: Nitin Gupta, Srinivasan Venkatachary, Lingkun Chu, Steven D. Baker

  Continuation: "heading hierarchy having two or more heading levels hierarchically arranged in a parent-child relationship." The parent-child chain is the Contextual Vector. Reference: Claim 1 of US-11409748-B1.

• US 9448992 B2 Natural language search results for intent queriesGoogle, 2013 · Active, term to 2033 · Inventors: Tomer Shmiel, Dvir Keysar, Yonatan Erez

  Heading-text pairs chained together form a progression. Reference: Claim 1.

• US 7913163 B1 Determining semantically distinct regions of a documentGoogle, 2004 · Expired 2024 · Inventors: Yonatan Zunger

  Chunk tree represents the logical progression as nested regions. Reference: Fig. 2.

• US 10354292 B1 Systems and methods for generating navigation filtersGoogle, 2016 · Active, term to 2036 · Inventors: Shreyas Doshi, Venky Ramachandran, Gaurav Garg, Poorva Hari Potdar (+3 more)

  Heading-based navigation filter extraction depends on a straight heading progression. Reference: Detailed description HTML example.

• US 9069855 B2 Modifying a hierarchical data structure according to pseudo-renderingGoogle, 2011 · Active, term to 2031 · Inventors: Yonatan Zunger

  Pseudo-rendering preserves document structure in the hierarchical representation, straight contextual vectors render cleanly. Reference: Fig. 2.

Google Content Warehouse leak 2 attributes

• QualityAuthorityTopicEmbeddingsVersionedItem.pageEmbedding the page-level embedding literally represents the straight-line topical vector; Contextual Drift = pageEmbedding deviating from the intended vector.
• QualityAuthorityTopicEmbeddingsVersionedItem.siteRadius page-vs-site embedding deviation. Drift at the page level shows up as inflated siteRadius.

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Patent evidence 2 patents

• US 9959315 B1 Context scoring adjustments for answer passagesGoogle, 2014 · Active, term to 2034 · Inventors: Nitin Gupta, Srinivasan Venkatachary, Lingkun Chu, Steven D. Baker

  Explicit heading-hierarchy diagram: "heading 310 is the root heading at the root level; headings 330 and 360 are child headings… at a first level… headings 334 and 338 are child headings of 330, and are at a second level." The exact H1-H2-H3-H4 model Koray requires. Reference: Detailed description Fig. 3 heading hierarchy.

• US 11409748 B1 Context scoring adjustments for answer passagesGoogle, 2018 · Active, term to 2038 · Inventors: Nitin Gupta, Srinivasan Venkatachary, Lingkun Chu, Steven D. Baker

  Continuation of US-9959315-B1 with claim 1 explicitly reciting the parent-child heading hierarchy. Reference: Claim 1.

Google Content Warehouse leak 3 attributes

• IndexingDocjoinerDataVersion.indexingMlPageStructure ML-based page structure annotation detects heading hierarchy.
• IndexingDocjoinerDataVersion.fontsAnalysisAnnotatorFontsAnnotation font-size and font-weight analysis, exactly the typography that Koray says Google reads to construct heading dependency trees.
• IndexingDocjoinerDataVersion.indexingAnnotationsWaPassagesQueryToPassageAnnotations passage-level mapping relies on heading structure.

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Patent evidence 1 patent

• US 11886533 B2 Transferable neural architecture for structured data extractionGoogle, 2020 · Active, term to 2040 · Inventors: Ying Sheng, Yuchen Lin, Sandeep Tata, Nguyen Vo

  Reads DOM nodes and extracts field-of-interest labels and node-pair relations. Algorithmic Authorship aligns prose output with this extractor's expectation. Reference: Claim 1 and Claim 9.

Google Content Warehouse leak 4 attributes

• IndexingDocjoinerDataVersion.repositoryProtiumGoldmineTriplificationResult the extractor that benefits most from Algorithmic Authorship. Clean S-P-O writing feeds this directly.
• QualityNsrPQData.contentEffort LLM-based effort scoring; strict Algorithmic Authorship rules show up as higher effort scores.
• IndexingDocjoinerDataVersion.qualityAuthorshipAuthorAnnotations explicit per-document authorship annotation pipeline.
• QualityCopiaFireflySiteSignal.latestBylineDateSec Firefly site signal reads byline authorship consistency.

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Patent evidence 4 patents

• US 9448992 B2 Natural language search results for intent queriesGoogle, 2013 · Active, term to 2033 · Inventors: Tomer Shmiel, Dvir Keysar, Yonatan Erez

  Parses heading-text pairs, reads H1 and First Subordinate Text as a joint IR Zone. Reference: Claim 1.

• US 9959315 B1 Context scoring adjustments for answer passagesGoogle, 2014 · Active, term to 2034 · Inventors: Nitin Gupta, Srinivasan Venkatachary, Lingkun Chu, Steven D. Baker

  Heading vector = H1 → H2 → H3 chain; each heading is its own IR Zone contributing to the vector. Reference: Claim 1.

• US 8914359 B2 Ranking documents with social tagsMicrosoft, 2008 · Active, term to 2028 · Inventors: Yauhen Shnitko, Dmitriy Meyerzon

  Granted version with the same multi-zone scoring.

• US 11886533 B2 Transferable neural architecture for structured data extractionGoogle, 2020 · Active, term to 2040 · Inventors: Ying Sheng, Yuchen Lin, Sandeep Tata, Nguyen Vo

  Node content + XPath → structured output; URL path (XPath encoded) is explicitly fed into the extractor alongside node content. The URL-slug IR Zone is literally a feature. Reference: Claim 9 (XPath vector encoding).

Google Content Warehouse leak 7 attributes

• QualityPreviewSnippetDocumentFeatures.experimentalTitleSalientTermsScore Title IR zone.
• QualityPreviewSnippetDocumentFeatures.metaBoostScore Meta description IR zone.
• QualityPreviewSnippetDocumentFeatures.schemaOrgDescriptionBoostScore Schema.org description IR zone.
• QualityPreviewSnippetDocumentFeatures.leadingtextDistanceScore Lead-text IR zone.
• QualityPreviewSnippetDocumentFeatures.salientPositionBoostScore Salient-position IR zone.
• WWWMetaTag.content and .name, raw meta tags.
• ImageExifImageEmbeddedMetadata.author + Image alt / caption attributes, image IR zone.

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Patent evidence 5 patents

• US 9753901 B1 Identifying important document content using geometries of page elementsGoogle, 2013 · Active, term to 2033 · Inventors: Yifan Xu, Xiaofeng Mi

  Direct hit. Abstract: "Systems and techniques are provided for detecting columns of an electronic page based on a render of the electronic page and identification of one or more columns based on the render. A column of interest may be identified based on [geometric properties]." The explicit Google patent that identifies the Centerpiece visually. Reference: Abstract and claim 1.

• US 12393768 B2 Layout-aware multimodal pretrainingGoogle, 2020 · Active, term to 2040 · Inventors: Mingyang Zhang, Cheng Li, Tao Chen, Spurthi Amba Hombaiah (+3 more)

  Granted version.

• US 10042829 B2 Measurement of visibility of overlay contentGoogle, 2016 · Active, term to 2036 · Inventors: Robert Durbin, Justin Lewis

  Granted version of US-2017371853-A1.

• US 7913163 B1 Determining semantically distinct regions of a documentGoogle, 2004 · Expired 2024 · Inventors: Yonatan Zunger

  Chunk tree isolates the primary chunk, the Centerpiece lives in the top node. Reference: Fig. 2.

• US 9069855 B2 Modifying hierarchical data structure according to pseudo-renderingGoogle, 2011 · Active, term to 2031 · Inventors: Yonatan Zunger

  Pseudo-rendering elevates top-of-page components. Reference: Fig. 2.

Google Content Warehouse leak 4 attributes

• IndexingDocjoinerDataVersion.indexingMlPageStructure the ML-based page-structure classifier identifies the primary container.
• IndexingDocjoinerDataVersion.boilerplateVisibleBytesEstimates boilerplate measurement; non-boilerplate above-the-fold = Center-Piece.
• CompositeDocAdditionalChecksums.NoTransientChecksum96 checksum of content excluding boilerplate; the stable content "center piece."
• QualityPreviewSnippetExperimentalFeatures.isLikelyHomepage homepage layout signal; homepage Center-Piece has distinct treatment.

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Patent evidence 2 patents

• US 12393768 B2 Layout-aware multimodal pretrainingGoogle, 2020 · Active, term to 2040 · Inventors: Mingyang Zhang, Cheng Li, Tao Chen, Spurthi Amba Hombaiah (+3 more)

  Granted version.

• US 11886533 B2 Transferable neural architecture for structured data extractionGoogle, 2020 · Active, term to 2040 · Inventors: Ying Sheng, Yuchen Lin, Sandeep Tata, Nguyen Vo

  Extracts labeled nodes from DOM including image elements with alt text; Visual Semantics (ALT text + file-name conventions) feeds structured extraction. Reference: Claim 1 step C.

Google Content Warehouse leak 6 attributes

• IndexingDocjoinerDataVersion.fontsAnalysisAnnotatorFontsAnnotation typographic/visual analysis.
• IndexingDocjoinerDataVersion.photosGeoAnnotation photo-level annotations.
• ImageExifIPTCMetadata.creator IPTC image authorship metadata.
• ImageData.contentFirstCrawlTime image content age for visual freshness.
• IndexingDocjoinerDataVersion.repositoryAnnotationsRdfaRichSnippetsApplication family (30+ attrs), rich-image/structured visual annotations.
• HtmlrenderWebkitHeadlessProtoImage headless-render image.

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Patent evidence 3 patents

• US 8825471 B2 Unsupervised extraction of factsGoogle, 2006 · Expired 2026 · Inventors: Jonathan T. Betz, Shubin Zhao

  Extracts attribute-value facts from documents and uses them as seed pairs, grounded facts are Safe Answers that propagate. Reference: Abstract.

• US 10354188 B2 Extracting facts from unstructured informationMicrosoft, 2016 · Active, term to 2036 · Inventors: Achraf Abdel Moneim Tawfik Chalabi, Ahmed Mohamed Emad Morsi Abdelbaki, Brandon Robert Anderson, Eslam Kamal Abdel-Aal Abdel-Reheem (+4 more)

  Pipeline for extracting facts from unstructured documents, the engine that captures Safe Answers from text. Reference: Abstract.

• US 11475319 B2 Extracting facts from unstructured informationMicrosoft, 2019 · Active, term to 2039 · Inventors: Achraf Abdel Maneim Tawfik Chalabi, Ahmed Mohamed Emad Morsi Abdelbaki, Brandon Robert Anderson, Eslam Kamal Abdel-Aal Abdel-Reheem (+4 more)

  Granted continuation.

Google Content Warehouse leak 3 attributes

• IndexingDocjoinerDataVersion.qualityProseJsonLdAnnotation JSON-LD annotation extracts the Safe Answer factual triple.
• IndexingDocjoinerDataVersion.repositoryProtiumGoldmineTriplificationResult the triple extractor validates Safe Answer facts.
• IndexingDocjoinerDataVersion.researchCoauthorOpinionAnnotations "coauthor opinion" annotations capture perspective shift.

Related research and notes

• Dong et al., Knowledge Vault (KDD 2014), probabilistic facts combined with per-source priors; the Safe Answer is the high-prior consensus.
• Attributed Question Answering (Bohnet et al.) (arXiv:2212.08037, 2022), grounds the attribution of a Perspective Shift to a specific source, separate from the Safe Answer.

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Patent evidence 6 patents

• US 8825471 B2 Unsupervised extraction of factsGoogle, 2006 · Expired 2026 · Inventors: Jonathan T. Betz, Shubin Zhao

  Attribute-value pairs are Points; the document region they live in is the Zone. Reference: Abstract.

• US 10354188 B2 Extracting facts from unstructured informationMicrosoft, 2016 · Active, term to 2036 · Inventors: Achraf Abdel Moneim Tawfik Chalabi, Ahmed Mohamed Emad Morsi Abdelbaki, Brandon Robert Anderson, Eslam Kamal Abdel-Aal Abdel-Reheem (+4 more)

  Pipeline distinguishes extraction region (Zone) from extracted fact (Point). Reference: Abstract.

• US 11475319 B2 Extracting facts from unstructured informationMicrosoft, 2019 · Active, term to 2039 · Inventors: Achraf Abdel Maneim Tawfik Chalabi, Ahmed Mohamed Emad Morsi Abdelbaki, Brandon Robert Anderson, Eslam Kamal Abdel-Aal Abdel-Reheem (+4 more)

  Granted continuation.

• US 11886533 B2 Transferable neural architecture for structured data extractionGoogle, 2020 · Active, term to 2040 · Inventors: Ying Sheng, Yuchen Lin, Sandeep Tata, Nguyen Vo

  Claim 1 generates node labels for first and second DOM nodes (Zone = node, Point = label). Claim 9 generates pair labels for node-pairs (relation Points). The explicit Zone-to-Point extraction. Reference: Claim 1 step E, Claim 9 pair-label step.

• US 7913163 B1 Determining semantically distinct regions of a documentGoogle, 2004 · Expired 2024 · Inventors: Yonatan Zunger

  The chunk-tree region is the Zone; the extracted content inside is the Point. Reference: Fig. 2.

• US 9959315 B1 Context scoring adjustments for answer passagesGoogle, 2014 · Active, term to 2034 · Inventors: Nitin Gupta, Srinivasan Venkatachary, Lingkun Chu, Steven D. Baker

  Heading-anchored Zones; answer passages are the Points. Reference: Claim 1.

Google Content Warehouse leak 3 attributes

• IndexingDocjoinerDataVersion.repositoryProtiumGoldmineTriplificationResult Information Points persist as triples here.
• IndexingDocjoinerDataVersion.indexingAnnotationsWaPassagesQueryToPassageAnnotations Information Zones persist as passages here.
• RepositoryWebrefEntityAnnotations.segmentMentions segments are the Information Zones; mentions inside are the Information Points.

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Patent evidence 4 patents

• US 9959315 B1 Context scoring adjustments for answer passagesGoogle, 2014 · Active, term to 2034 · Inventors: Nitin Gupta, Srinivasan Venkatachary, Lingkun Chu, Steven D. Baker

  Direct hit. Claim 1 and dependent claims: "heading vector includes text of each heading in the path in the heading hierarchy from the root heading to the respective heading" and "determining a context score based, at least in part, on the heading vector comprises determining a plurality of match scores." The exact Heading Vector mechanism. Reference: Dependent claim on match-score computation.

• US 11409748 B1 Context scoring adjustments for answer passagesGoogle, 2018 · Active, term to 2038 · Inventors: Nitin Gupta, Srinivasan Venkatachary, Lingkun Chu, Steven D. Baker

  Continuation. Claim 1 recites the heading-hierarchy traversal. Reference: Claim 1.

• US 9448992 B2 Natural language search results for intent queriesGoogle, 2013 · Active, term to 2033 · Inventors: Tomer Shmiel, Dvir Keysar, Yonatan Erez

  Heading-text pairs across a document, a drifted H2's text pair would not match the primary intent template. Reference: Claim 1.

• US 7913163 B1 Determining semantically distinct regions of a documentGoogle, 2004 · Expired 2024 · Inventors: Yonatan Zunger

  Chunk-tree segmentation isolates each section; drift becomes measurable as an outlier chunk. Reference: Fig. 2.

Google Content Warehouse leak 3 attributes

• IndexingDocjoinerDataVersion.indexingMlPageStructure ML page structure annotates heading vector.
• QualityAuthorityTopicEmbeddingsVersionedItem.siteRadius + .pageEmbedding, Contextual Drift shows as page_embedding deviation from the intended Macro Context.
• QualityNsrNsrData.siteQualityStddev "Estimate of site's PQ rating stddev-spread of the page-level PQ ratings of a site." Drift at scale raises stddev.

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Patent evidence 3 patents

• US 10430445 B2 Text indexing and passage retrievalMicrosoft, 2014 · Active, term to 2034 · Inventors: Richard S. Crouch, Marisa F. Boston, Ali Erdem Ozcan, Peter R. Stubley

  Granted version of US-2016078102-A1.

• US 9959315 B1 Context scoring adjustments for answer passagesGoogle, 2014 · Active, term to 2034 · Inventors: Nitin Gupta, Srinivasan Venkatachary, Lingkun Chu, Steven D. Baker

  Scores answer passages, the passage-level unit that rewards dense factual sentences. Reference: Claim 1.

• US 9323827 B2 Identifying key terms related to similar passagesGoogle, 2008 · Active, term to 2028 · Inventors: William N. Schilit, Okan Kolak

  Extracts pre-context and post-context within a threshold number of words, defines a sliding window around each passage. Reference: Claim 1 (pre-context/post-context thresholds).

Google Content Warehouse leak 4 attributes

• IndexingDocjoinerDataVersion.indexingAnnotationsWaPassagesQueryToPassageAnnotations passage-level chunking is Google's own sliding window.
• IndexingDocjoinerDataVersion.boilerplateVisibleBytesEstimates "visible bytes" measurement feeds the per-byte density calculation.
• DocProperties.numTokens token count against which density is computed.
• IndexingDocjoinerDataVersion.repositoryProtiumGoldmineTriplificationResult triples per byte = pure Information Density.

Related research and notes

• REALM (Guu et al., ICML 2020, arXiv:2002.08909). Retrieves passages for augmented generation; dense passages extracted preferentially.

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Patent evidence 4 patents

• US 9753901 B1 Identifying important document content using geometries of page elementsGoogle, 2013 · Active, term to 2033 · Inventors: Yifan Xu, Xiaofeng Mi

  Direct hit on Web Layout Geometry. Identifies important content from the render based on geometric column analysis. Reference: Abstract and claim 1.

• US 12393768 B2 Layout-aware multimodal pretrainingGoogle, 2020 · Active, term to 2040 · Inventors: Mingyang Zhang, Cheng Li, Tao Chen, Spurthi Amba Hombaiah (+3 more)

  Granted version.

• US 7913163 B1 Determining semantically distinct regions of a documentGoogle, 2004 · Expired 2024 · Inventors: Yonatan Zunger

  Chunk-tree per visual region, each region is a component in Koray's dictionary. Reference: Fig. 2.

• US 9069855 B2 Modifying hierarchical data structure according to pseudo-renderingGoogle, 2011 · Active, term to 2031 · Inventors: Yonatan Zunger

  Pseudo-rendering preserves component geometry during indexing. Reference: Fig. 2.

Google Content Warehouse leak 4 attributes

• IndexingDocjoinerDataVersion.indexingMlPageStructure ML page structure identifies layout components.
• GoogleCloudDocumentaiV1DocumentPage.layout and GoogleCloudDocumentaiV1DocumentPageBlock.layout, explicit layout objects.
• GoogleCloudDocumentaiV1DocumentPageAnchorPageRef.boundingPoly bounding polygon for layout elements; geometric component storage.
• IndexingDocjoinerDataVersion.fontsAnalysisAnnotatorFontsAnnotation font-level component dictionary.

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Patent evidence 2 patents

• US 11829435 B2 Secure digital assistant integration in web pagesGoogle, 2022 · Active, term to 2042 · Inventors: Adam Coimbra, Nikhil Rao, Hao Wu, Ian Baker (+2 more)

  Direct patent on page-side action/intent payloads consumed by digital assistants, the server-side mechanism that distinguishes assistant consumers from human readers. Reference: Example action JSON.

• US 11086959 B2 Secure digital assistant integration in web pagesGoogle, 2020 · Active, term to 2040 · Inventors: Adam Coimbra, Nikhil Rao, Hao Wu, Ian Baker (+2 more)

  Granted version with the same assistant-targeted payload mechanism.

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Patent evidence 2 patents

• US 10354188 B2 Extracting facts from unstructured informationMicrosoft, 2016 · Active, term to 2036 · Inventors: Achraf Abdel Moneim Tawfik Chalabi, Ahmed Mohamed Emad Morsi Abdelbaki, Brandon Robert Anderson, Eslam Kamal Abdel-Aal Abdel-Reheem (+4 more)

  Pipeline for extracting facts from unstructured text, reads the Unstructured Factual quadrant into the Structured Factual output. Reference: Abstract.

• US 10255380 B2 Identifying and structuring related dataMicrosoft, 2013 · Active, term to 2033 · Inventors: Jun Yan, Ning Liu, Lei Ji, Zheng Chen

  Claim 1: receives input entity and displays related attributes in a structured UI region, Structured Factual quadrant automation. Reference: Claim 1.

Google Content Warehouse leak 4 attributes

• IndexingDocjoinerDataVersion.qualityProseJsonLdAnnotation JSON-LD represents the structured factual quadrant.
• IndexingDocjoinerDataVersion.researchCoauthorOpinionAnnotations opinionated-content annotation.
• IndexingDocjoinerDataVersion.repositoryAnnotationsListSummaryAnnotations list-summary (structured) annotation.
• CompressedQualitySignals.productReviewPReviewPage and .productReviewPUhqPage, explicit review-page scoring (opinionated/structured quadrant).

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Patent evidence 6 patents

• US 11086959 B2 Secure digital assistant integration in web pagesGoogle, 2020 · Active, term to 2040 · Inventors: Adam Coimbra, Nikhil Rao, Hao Wu, Ian Baker (+2 more)

  JSON "action": [{ "intentName": "actions.intent.ORDER_RIDE", "fulfillment": [{"urlTemplate": ...}]}], Google's exact action-trigger payload format on pages. Information Cards and Action Triggers are formally defined here. Reference: Example JSON payload.

• US 11829435 B2 Secure digital assistant integration in web pagesGoogle, 2022 · Active, term to 2042 · Inventors: Adam Coimbra, Nikhil Rao, Hao Wu, Ian Baker (+2 more)

  Continuation with the same action-trigger JSON.

• US 9753901 B1 Identifying important document content using geometries of page elementsGoogle, 2013 · Active, term to 2033 · Inventors: Yifan Xu, Xiaofeng Mi

  Geometric detection of important content, Information Cards occupy geometrically distinguishable blocks. Reference: Abstract.

• US 10354292 B1 Generating navigation filtersGoogle, 2016 · Active, term to 2036 · Inventors: Shreyas Doshi, Venky Ramachandran, Gaurav Garg, Poorva Hari Potdar (+3 more)

  Detailed description HTML example recognizes heading-anchored card blocks (<h2>Featured Brands</h2> + list items as a card). Reference: Detailed description.

• US 10255380 B2 Identifying and structuring related dataMicrosoft, 2013 · Active, term to 2033 · Inventors: Jun Yan, Ning Liu, Lei Ji, Zheng Chen

  Displays related attributes in a structured first region, the card formalism. Reference: Claim 1.

• US 9158849 B2 Synthesis of webpage snippets using sub-pagesGoogle, 2013 · Active, term to 2033 · Inventors: Junli Xian, Harvey Jones, Samarth Keshava

  Sub-page cards aggregated into a snippet synthesis. Reference: Claim 1.

Google Content Warehouse leak 3 attributes

• IndexingDocjoinerDataVersion.repositoryAnnotationsRdfaRdfaRichSnippetsApplication the Rich Snippets Application proto with 30+ fields (rating, screenUrlThumbnail, applicationUrl, price, inAppPurchase, supportsChromecast) is exactly the "structured information card" extraction target.
• IndexingDocjoinerDataVersion.richsnippetsPageMap and .richsnippetsDataObject, PageMap-level rich-snippet storage.
• IndexingDocjoinerDataVersion.richsnippetsSchemaThing schema.org Thing-type storage.

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Patent evidence 1 patent

• US 8868402 B2 Construction of text classifiersGoogle, 2009 · Active, term to 2029 · Inventors: Dmitry Korolev, Hartmut Maennel, Matthias Heiler, Michael Schaer (+3 more)

  n-gram candidate phrases per topic, the n-gram co-occurrence substrate for distributional topic classification. Reference: Detailed description.

Google Content Warehouse leak 3 attributes

• QualityAuthorityTopicEmbeddingsVersionedItem.siteEmbedding and .pageEmbedding, compressed vector representations built from distributional statistics.
• QualityNsrNsrData.site2vecEmbeddingEncoded "site2vec" is a distributional-semantic site vector.
• CompressedQualitySignals.topicEmbeddingsVersionedData versioned topic embeddings in the compressed quality signal set.

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Google Content Warehouse leak 4 attributes

• IndexingDocjoinerDataVersion.repositoryProtiumGoldmineTriplificationResult sentence-level S-P-O extraction.
• IndexingDocjoinerDataVersion.repositoryProtiumGoldmineTypeAssertionOutput type assertions extracted from sentence structure.
• IndexingDocjoinerDataVersion.nlpSaftNlxDocAnnotatorMigration explicit NLP SAFT/NLX annotator pipeline for sentence parsing.
• NlpSaftDocument.bylineDate .contentage, SAFT document annotations.

Related research and notes

• Semantic Role Labeling (SRL) survey (Palmer, Gildea, Xue, 2010). The academic foundation.

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Patent evidence 2 patents

• US 8407204 B2 Minimizing visibility of stale content including revising crawl intervalsGoogle, 2011 · Active, term to 2031 · Inventors: Anton P. T. Carver

  P_stale, PR_pagerank, T_click_rate, T_content_update_rate formula balances staleness probability against retrieval cost. Reference: Detailed description formula.

• US 8645342 B2 Scope based indexing prioritizationMicrosoft, 2012 · Active, term to 2032 · Inventors: Michael J. Novak, Christopher C. McConnell

  Scope-based indexing prioritization is cost-aware. Reference: Abstract.

Google Content Warehouse leak 4 attributes

• CrawlerChangerateUrlChange.additionalChangesMerged "Duplicate UrlChanges crawled within a specified time range will be merged together." Deduplication reduces retrieval cost.
• CompositeDocAdditionalChecksums.NoTransientChecksum96 stripped-boilerplate checksum; cheaper to compare.
• IndexingDocjoinerDataVersion.boilerplateVisibleBytesEstimates explicit boilerplate byte accounting.
• PerDocData.hostAge "used in twiddler to sandbox fresh spam in serving time." Hosts with low host age have higher retrieval cost (serving-time sandboxing).

Related research and notes

• Barroso, Dean, Hölzle, "Web Search for a Planet: The Google Cluster Architecture" (IEEE Micro 2003). Grounds cost-of-retrieval economics at Google scale.

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Patent evidence 1 patent

• US 12314275 B2 Cross-list learning to rankGoogle, 2023 · Active, term to 2043 · Inventors: Gil Shamir

  "Cross-List Learning to Rank", modern re-ranking across candidate lists. Reference: Subheading section.

Google Content Warehouse leak 5 attributes

• CompressedQualitySignals.experimentalQstarSignal .experimentalQstarSiteSignal, .experimentalQstarDeltaSignal, Q* experimental signals populated at serving time for re-ranking.
• CompositeDoc.docinfoPassthroughAttachments "used in post-doc twiddlers," direct storage for re-ranking signals.
• PerDocData.datesInfo "Used in FreshnessTwiddler." Re-ranking by freshness.
• IndexingDupsLocalizedLocalizedCluster.boostSourceBlocker "a result which can be a boost target but should itself not be boosted." Re-ranking twiddler control.
• IndexingDocjoinerDataVersion.qualityTwiddlerDomainClassification domain classifier used in Twiddler re-ranking.

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Patent evidence 4 patents

• US 8527523 B1 Expert-based computerized analysis of digital documentsMicrosoft, 2009 · Active, term to 2029 · Inventors: Yiftach Ravid

  Graduated rankings allow early culling of low-priority documents, predictive IR culling. Reference: Detailed description.

• US 8706742 B1 Expert-based computerized analysis of digital documentsMicrosoft, 2012 · Active, term to 2032 · Inventors: Yiftach Ravid, Liad Tal-Rothschild

  Continuation.

• US 9959315 B1 Context scoring adjustments for answer passagesGoogle, 2014 · Active, term to 2034 · Inventors: Nitin Gupta, Srinivasan Venkatachary, Lingkun Chu, Steven D. Baker

  Heading vectors scored first; passages with low heading-vector match don't advance to deeper scoring. Reference: Claim 1.

• US 9448992 B2 Natural language search results for intent queriesGoogle, 2013 · Active, term to 2033 · Inventors: Tomer Shmiel, Dvir Keysar, Yonatan Erez

  Heading-text pairs used to classify documents against intent templates before full content scoring. Reference: Claim 1.

Google Content Warehouse leak 4 attributes

• QualityPreviewSnippetDocumentFeatures.leadingtextDistanceScore direct measurement of where the useful answer is relative to the top. If distance is high, Predictive IR skips the rest.
• QualityPreviewSnippetDocumentFeatures.salientPositionBoostScore salient-term proximity to the top; predictive input.
• QualityPreviewSnippetBrainFeatures.modelScore SnippetBrain is Google's own predictive IR model for snippet selection.
• QualityPreviewSnippetDocumentFeatures.unstableTokensScore unstable tokens near top signal unpredictable content; Predictive IR penalizes.

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Google Content Warehouse leak 4 attributes

• IndexingDocjoinerDataVersion.repositoryProtiumGoldmineTriplificationResult the actual Google triple extractor. OpenIE is executed here.
• IndexingDocjoinerDataVersion.repositoryProtiumGoldmineTypeAssertionOutput type assertion extraction, part of OpenIE output.
• IndexingDocjoinerDataVersion.repositoryAnnotationsSporeSporeEtlAnnotation and .repositoryAnnotationsSporeExtractorTripleAnnotations, Spore ETL and triple extraction, feeding KG.
• IndexingDocjoinerDataVersion.websitetoolsFeedsSporeFeedsSporeTriples Spore triples storage.

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Google Content Warehouse leak 4 attributes

• IndexingDocjoinerDataVersion.nlpSaftNlxDocAnnotatorMigration SAFT (Structured Annotation Framework for Text) is Google's SRL-capable NLP pipeline.
• NlxDataSchemaMention.kind "Mention kind, typically 'referential', 'attributive', or 'modifier'." Direct SRL role encoding.
• NlxDataSchemaMention.type "Mention type, typically 'named' (for name mentions) or 'nominal'. More types include 'pronominal', 'conjoined' for conjoined mention construction, and 'non-referential' for non-referential pronoun mentions." Frame-semantic typing.
• NlxDataSchemaTokenDependencyEdge token dependency edge; the SRL arc store.

Related research and notes

• PropBank (Palmer et al., 2005), predicate-argument annotation standard.
• SRL survey (Palmer, Gildea, Xue, 2010).

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Google Content Warehouse leak 3 attributes

• IndexingDocjoinerDataVersion.qualitySalientTermsQuerySmearingModel "query smearing" expands queries via lexical relations (synonyms, hypernyms).
• IndexingDocjoinerDataVersion.qualitySalientTermsSalientTermSet salient term sets include lexical variants.
• GoogleCloudContentwarehouseV1SynonymSet + .ListSynonymSetsResponse + SynonymSetSynonym, explicit synonym-set storage in Cloud Content Warehouse API.

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Patent evidence 6 patents

• US 10387437 B2 Query rewriting using session informationGoogle, 2017 · Active, term to 2037 · Inventors: Marcin M. Nowak-Przygodzki, Behshad Behzadi

  Direct Google patent on session-informed query rewriting. Method: receives a search query during a user session; obtains prior search queries by the user; generates candidate rewrites. Reference: Claim 1.

• US 9547690 B2 Query rewriting using session informationGoogle, 2014 · Active, term to 2034 · Inventors: Marcin M. Nowak-Przygodzki, Behshad Behzadi

  Earlier grant with the same mechanism.

• US 9165033 B1 Efficient query rewritingGoogle, 2014 · Active, term to 2034 · Inventors: Paul A. Tucker, Amit Singhal, Eric Jackson

  Selects an original query from records of previously received user queries and automatically rewrites, directly validated with user response. Reference: Claim 1.

• US 11960554 B2 Unsupervised entity and intent identification for improved search query relevanceMicrosoft, 2021 · Active, term to 2041 · Inventors: Gautam Prasad, Varun Appaswami, Bhanu Teja Chunduri

  Extracts entity + intent from search-page data, reads query semantics behaviorally. Reference: Abstract.

• US 11068554 B2 Unsupervised entity and intent identificationMicrosoft, 2019 · Active, term to 2039 · Inventors: Gautam Prasad, Varun Appaswami, Bhanu Teja Chunduri

  Granted continuation.

• US 9870356 B2 Techniques for inferring the unknown intents of linguistic itemsMicrosoft, 2014 · Active, term to 2034 · Inventors: Dilek Hakkani-Tür, Fethiye Asli Celikyilmaz, Larry P. Heck, Gokhan Tur (+1 more)

  Deterministic intent assignment on queries. Reference: Abstract.

Google Content Warehouse leak 3 attributes

• IndexingDocjoinerDataVersion.qualitySalientTermsQuerySmearingModel query smearing is Google's canonical-query mapping.
• IndexingDocjoinerDataVersion.qualityNavboostCrapsCrapsPatternSignal click-pattern signals that validate "buy" and "sell" being Markovian-equivalent for commercial intent.
• KnowledgeAnswersIntentQueryLocalSignals.minSalientTermSimilarity minimum salient-term similarity between results; the query-similarity threshold.

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Patent evidence 1 patent

• US 8868402 B2 Construction of text classifiersGoogle, 2009 · Active, term to 2029 · Inventors: Dmitry Korolev, Hartmut Maennel, Matthias Heiler, Michael Schaer (+3 more)

  N-gram variety across a corpus is the base of vocabulary richness. Reference: Detailed description.

Google Content Warehouse leak 3 attributes

• RepositoryWebrefAnnotatorProfile.numTokens .numEntities, .numMentions, .numCandidateMentions, raw counts; richness = unique-terms-per-doc.
• QualityNsrPQData.contentEffort content effort implicitly correlates with vocabulary richness.
• DocProperties.numTokens token count baseline.

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Patent evidence 6 patents

• US 8140449 B1 Detecting novel document contentGoogle, 2008 · Active, term to 2028 · Inventors: M. Bharath Kumar, Krishna Bharat

  Claim 1: "identifying textual sequences; identifying based on the one or more textual sequences a presence of novel content in the document." Direct Google patent on novel-content detection, the engineering substrate of Information Gain. Reference: Claim 1.

• US 7451120 B1 Detecting novel document contentGoogle, 2006 · Expired 2026 · Inventors: M. Bharath Kumar, Krishna Bharat

  Earlier grant in the novel-content family.

• US 8521665 B1 Detecting novel document contentGoogle, 2012 · Active, term to 2032 · Inventors: M. Bharath Kumar, Krishna Bharat

  Continuation.

• US 10970353 B1 Ranking content using content and content authorsGoogle, 2019 · Active, term to 2039 · Inventors: Douwe Osinga, Stefan Christoph

  Background: "determining the originality of content", direct successor framing. Reference: Background.

• US 7984054 B2 Representative document selection for sets of duplicate documentsGoogle, 2009 · Active, term to 2029 · Inventors: Daniel Dulitz, Alexandre A. Verstak, Sanjay Ghemawat, Jeffrey A. Dean

  Deduplicates and keeps the representative; Information Gain is the measure that distinguishes a new document from the representative. Reference: Claim 1.

• US 8280861 B1 Identifying duplicate electronic content based on metadataGoogle, 2011 · Active, term to 2031 · Inventors: Tim Park, Dmitry Dolinsky

  Hash-based duplicate identification. Reference: Claim 1.

Google Content Warehouse leak 4 attributes

• QualityNsrPQData.contentEffort LLM-based effort estimation; unique facts raise effort score.
• CompressedQualitySignals.productReviewPUhqPage "The possibility of a page being a high quality review page." Unique data / testing elevates this.
• IndexingDocjoinerDataVersion.repositoryProtiumGoldmineTriplificationResult triple count = fact density = raw information gain proxy.
• IndexingDocjoinerDataVersion.researchStructuredDataKnowledgeExplorationAnnotations structured-data knowledge-exploration annotation evaluates unique knowledge contribution.

Related research and notes

• Google patent (not retrieved here but canonical): US-11354316-B2, Systems and methods for contextual estimation of link information gain (Google). Explicit "link information gain" terminology Koray cites.
• Attributed Question Answering (Bohnet et al., arXiv:2212.08037). Grounds Information Gain at the attribution-to-source level.

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Patent evidence 2 patents

• US 8825471 B2 Unsupervised extraction of factsGoogle, 2006 · Expired 2026 · Inventors: Jonathan T. Betz, Shubin Zhao

  Seed attribute-value propagation across sources, truth-range aggregation at scale. Reference: Abstract.

• US 10354188 B2 / US 11475319 B2 Extracting facts from unstructured informationMicrosoft, 2016/2019 · Active, term to 2036 · Inventors: Achraf Abdel Moneim Tawfik Chalabi, Ahmed Mohamed Emad Morsi Abdelbaki, Brandon Robert Anderson, Eslam Kamal Abdel-Aal Abdel-Reheem (+4 more)

  Facts extracted from unstructured sources, provides the sample population for consensus. Reference: Abstract.

Google Content Warehouse leak 4 attributes

• IndexingDocjoinerDataVersion.repositoryAnnotationsSporeAuthoritySignal Spore's authority signal reconciles source consensus.
• GeostoreRankSignalProto.rank "A value in the range [0, 1] estimating Oyster Rank." Rank is a consensus-reconciled score.
• GeostoreTrustSignalsProto.sourceTrust + GeostoreSourceTrustProto.level, source trust hierarchy for conflict reconciliation.
• RepositoryWebrefAggregatedEntityNameScores.entityScore aggregated-across-web entity-name scores; consensus-level.

Related research and notes

• Knowledge-Based Trust (KBT) (Dong et al., VLDB 2015, arXiv:1502.03519). Estimates the trustworthiness of web sources by factual-agreement rates with a KG.
• Knowledge Vault (Dong et al., KDD 2014). Probabilistic fact fusion across many sources.
• PaLM 2 fact verification / MinWMD (various 2023-2024). Modern LLM-era truth-range estimators.

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Patent evidence 5 patents

• US 10387437 B2 Query rewriting using session informationGoogle, 2017 · Active, term to 2037 · Inventors: Marcin M. Nowak-Przygodzki, Behshad Behzadi

  Session-informed query rewriting generates candidate rewrites, fan-out. Reference: Claim 1.

• US 9165033 B1 Efficient query rewritingGoogle, 2014 · Active, term to 2034 · Inventors: Paul A. Tucker, Amit Singhal, Eric Jackson

  Rewrites selected from previously received user queries. Reference: Claim 1.

• US 12530412 B2 Search-query suggestions using reinforcement learningMicrosoft, 2021 · Active, term to 2041 · Inventors: Praveen Kumar Bodigutla

  Granted version.

• US 7529736 B2 / US 2006253428 A1 Performant relevance improvements in search query resultsMicrosoft, 2005 · Expired 2025 · Inventors: Sanjeev Katariya, Qi Yao, Jun Liu, Adwait Ratnaparkhi (+1 more)

  Implicit user search query feedback → heuristic query rewrites. Reference: Claim 8.

• US 11983173 B2 Schema-agnostic query templatesGoogle, 2019 · Active, term to 2039 · Inventors: Anastasios Kementsietsidis, Jay Yogeshbhai Pandya, Tingting TANG, Laurren Kanner

  Templates generate many surface queries over an entity class, the "synthetic queries" counterpart. Reference: Abstract.

Google Content Warehouse leak 4 attributes

• IndexingDocjoinerDataVersion.qualitySalientTermsQuerySmearingModel query smearing/augmentation.
• IndexingDocjoinerDataVersion.indexingAnnotationsWaPassagesQueryToPassageAnnotations query-to-passage mapping includes synthetic queries generated from passage text.
• IndexingDocjoinerDataVersion.indexingAnnotationsWaPassagesTrimmedQueryToPassageAnnotations trimmed synthetic query index.
• QualityPreviewSnippetDocumentFeatures.salientTermsScore synthetic queries derive from salient terms.

Related research and notes

• Query fan-out in AI Overviews / Gemini, Koray cites Google's AI Mode fan-out directly; the engineering foundation is in US-10387437-B2.

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Patent evidence 2 patents

• US 11829435 B2 Secure digital assistant integration in web pagesGoogle, 2022 · Active, term to 2042 · Inventors: Adam Coimbra, Nikhil Rao, Hao Wu, Ian Baker (+2 more)

  Action JSON payloads consumed by assistants, assistant-side consumption gate. Reference: Example JSON.

• US 11886533 B2 Transferable neural architecture for structured data extractionGoogle, 2020 · Active, term to 2040 · Inventors: Ying Sheng, Yuchen Lin, Sandeep Tata, Nguyen Vo

  Low-cost DOM extraction is the structural prerequisite that lets a page pass an LLM's recruitment gate. Reference: Claim 1.

Google Content Warehouse leak 5 attributes

• QualityPreviewSnippetDocumentFeatures.leadingtextDistanceScore the recruitment-gate signal: how fast the answer appears.
• QualityPreviewSnippetBrainFeatures.modelScore SnippetBrain pre-filters candidates (gate).
• QualityNsrPQData.contentEffort low-effort docs discarded at the gate.
• CompressedQualitySignals.lowQuality "S2V low quality score … applied in Qstar." Low-quality gate.
• IndexingDocjoinerDataVersion.spamScamUnauthoritativeSite unauthoritative-site exclusion gate.

Related research and notes

• FreshLLMs (Vu et al., arXiv:2310.03214, 2023). Recency-filter behaviors in LLM retrieval, direct Agent Retrieval Bias study.
• REALM (Guu et al., ICML 2020, arXiv:2002.08909). Top-k retrieval gate before generation.
• Attributed Question Answering (Bohnet et al., arXiv:2212.08037, 2022). Grounds attribution-dependent gating.

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Google Content Warehouse leak 3 attributes

• IndexingDocjoinerDataVersion.repositoryProtiumGoldmineTypeAssertionOutput "type assertion" extraction. The Is-A formula directly produces this output.
• IndexingDocjoinerDataVersion.repositoryProtiumGoldmineTriplificationResult triple extraction captures the Is-A pattern as [Entity, is-a, SuperordinateEntity].
• IndexingDocjoinerDataVersion.commonsenseQueriosityGoldmineQnaAnnotations Q&A annotations rely on definitional sentence structure.

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Patent evidence 5 patents

• US 8782032 B2 Minimizing visibility of stale content including revising crawl intervalsGoogle, 2013 · Active, term to 2033 · Inventors: Anton P. T. Carver

  Detailed description: when documents change materially, scheduler sends notice to content indexer which replaces index entries. Tier-One documents enjoy higher re-crawl priority through this mechanic. Reference: Detailed description §308/§310.

• US 7987172 B1 Minimizing visibility of stale contentGoogle, 2004 · Expired 2024 · Inventors: Anton P. T. Carver

  Earliest grant in the family.

• US 8407204 B2 Minimizing visibility of stale contentGoogle, 2011 · Active, term to 2031 · Inventors: Anton P. T. Carver

  Continuation. Formula balancing staleness probability × PageRank × click rate × content update rate. Reference: Detailed description formula.

• US 8645342 B2 Scope based indexing prioritizationMicrosoft, 2012 · Active, term to 2032 · Inventors: Michael J. Novak, Christopher C. McConnell

  Priority ordering of indexing work by scope. Reference: Abstract.

• US 8296276 B2 Scope based indexing prioritizationMicrosoft, 2008 · Active, term to 2028 · Inventors: Michael J. Novak, Christopher C. McConnell

  Earlier grant.

Google Content Warehouse leak 4 attributes

• AnchorsAnchor.sourceType "TYPE_HIGH_QUALITY are from base documents. TYPE_MEDIUM_QUALITY … supplemental tier documents. TYPE_LOW_QUALITY … blackhole documents." Direct 3-tier evidence.
• AnchorsAnchor.sourceType also lists TYPE_FRESHDOCS, freshdocs is a separate serving tier with high-quality anchor equivalence.
• CompositeDocIndexingInfo.freshdocsCorpora "freshdocs corpora" tier annotation.
• IndexingDocjoinerDataVersion.indexingSelectionFreshdocsPassthroughExtension explicit freshdocs tier pass-through.

Related research and notes

• Reference: Pandu Nayak DOJ testimony (2023 US v. Google antitrust), public disclosure of Google's three-tier serving infrastructure (RAM / SSD / HDD tiers) Koray cites. Not a patent; evidentiary.

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Patent evidence 2 patents

• US 8407204 B2 Minimizing visibility of stale contentGoogle, 2011 · Active, term to 2031 · Inventors: Anton P. T. Carver

  Stale-content formula down-weights pages, site-wide decay produces a Negative state. Reference: Formula.

• US 7346839 B2 Information retrieval based on historical dataGoogle, 2003 · Expired 2023 · Inventors: Anurag Acharya, Matt Cutts, Jeffrey Dean, Paul Haahr (+6 more)

  Historical signals in positive direction = Positive state; negative = Negative state. Reference: Claim 1 and claims 10-38.

Google Content Warehouse leak 5 attributes

• CompressedQualitySignals.babyPandaDemotion "baby_panda_demotion: converted from QualityBoost.rendered.boost." Explicit Panda-style demotion persistence (= negative ranking state).
• CompressedQualitySignals.babyPandaV2Demotion "New BabyPanda demotion, applied on top of Panda."
• CompressedQualitySignals.pandaDemotion "This is the encoding of Panda fields in the proto SiteQualityFeatures."
• CompressedQualitySignals.authorityPromotion opposite direction: positive-state boost.
• IndexingDocjoinerDataVersion.indexingInstantBaseReindexWhitelistData base-reindex whitelist suggests state-change triggers reindex.

Related research and notes

• Content Warehouse API leak (May 2024): siteAuthority, siteQualityScore properties, the leaked attribute names Koray references for state classification.

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Patent evidence 4 patents

• US 8645342 B2 Scope based indexing prioritizationMicrosoft, 2012 · Active, term to 2032 · Inventors: Michael J. Novak, Christopher C. McConnell

  Scope-based inclusion prioritization. Reference: Abstract.

• US 8296276 B2 Scope based indexing prioritizationMicrosoft, 2008 · Active, term to 2028 · Inventors: Michael J. Novak, Christopher C. McConnell

  Earlier grant.

• US 9081861 B2 Uniform resource locator canonicalizationGoogle, 2008 · Active, term to 2028 · Inventors: Matthew Lloyd, Thomas Bergan

  Granted version.

• US 7984054 B2 Representative document selection for sets of duplicate documentsGoogle, 2009 · Active, term to 2029 · Inventors: Daniel Dulitz, Alexandre A. Verstak, Sanjay Ghemawat, Jeffrey A. Dean

  Collapses duplicates into one representative. Reference: Claim 1.

Google Content Warehouse leak 4 attributes

• QualityNsrNsrData.nsrdataFromFallbackPatternKey "data is coming from an average of other host chunks." Low-worth URLs fall back to host average, dragging the ratio.
• PerDocData.hostNsr host NSR score summarizes site-wide index worth.
• CompositeDoc.csePagerankCutoff "URL should only be selected for CSE Index if it's pagerank is higher than cse_pagerank_cutoff." PageRank cutoff gates index inclusion.
• IndexingDocjoinerDataVersion.qualityNsrSiteAuthority site-authority annotation drops when ratio tanks.

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Patent evidence 7 patents

• US 7346839 B2 Information retrieval based on historical dataGoogle, 2003 · Expired 2023 · Inventors: Anurag Acharya, Matt Cutts, Jeffrey Dean, Paul Haahr (+6 more)

  Scoring factors include document age, content update rate, anchor-text update rate, stale documents decay. Reference: Claim 1 and claims 10-38.

• US 8782032 B2 Minimizing visibility of stale contentGoogle, 2013 · Active, term to 2033 · Inventors: Anton P. T. Carver

  Scheduler monitors material changes and triggers re-indexing, the mechanism rewarding periodic re-configuration. Reference: Detailed description §308/§310.

• US 7987172 B1 Minimizing visibility of stale contentGoogle, 2004 · Expired 2024 · Inventors: Anton P. T. Carver

  Earlier grant.

• US 8407204 B2 Minimizing visibility of stale contentGoogle, 2011 · Active, term to 2031 · Inventors: Anton P. T. Carver

  Formula: P_stale, PR_pagerank, T_click_rate, T_content_update_rate. Higher update rate shortens the crawl interval and signals freshness. Reference: Formula.

• US 8549014 B2 Document scoring based on document content updateGoogle, 2011 · Active, term to 2031 · Inventors: Anurag Acharya, Jeffrey Dean, Paul Haahr, Monika Henzinger (+3 more)

  Score based on anchor-text and content update rate. Reference: Claim 1.

• US 9361369 B1 Clustering news online content based on content freshnessGoogle, 2012 · Active, term to 2032 · Inventors: Krishna Bharat, Michael Curtiss, Michael Schmitt

  Freshness score per content item. Reference: Claim 1.

• US 8407231 B2 Document scoring based on link-based criteriaGoogle, 2010 · Active, term to 2030 · Inventors: Anurag Acharya, Matt Cutts, Jeffrey Dean, Paul Haahr (+4 more)

  Appearance date per link; cadence tracking. Reference: Claim 1.

Google Content Warehouse leak 4 attributes

• CompositeDocQualitySignals.lastSignificantUpdate "Date Last Modified." The 6-month update rule aims to keep this fresh.
• IndexingDocjoinerDataVersion.qualityHistoricalContentAge historical content age tracking; aging content lose trust.
• QualityCopiaFireflySiteSignal.latestBylineDateSec site-level latest byline date; freshness of configuration.
• QualityNsrNsrData.nsrEpoch "The epoch from which this NSR value is coming from." Configuration must align to current epoch.

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Patent evidence 4 patents

• US 7296016 B1 Point-of-view searchingGoogle, 2003 · Expired 2023 · Inventors: Martin Farach-Colton, Monika H. Henzinger, Bay-Wei Chang

  Rank flows from designated example URLs. Quality Nodes are the example URLs. Reference: Detailed description.

• US 9165040 B1 Producing a ranking for pages using distances in a web-link graphGoogle, 2006 · Expired 2026 · Inventors: Nissan Hajaj

  Quality Nodes on the homepage are seed-close and thus boost adjacent pages. Reference: Claim 1.

• US 9158849 B2 Synthesis of webpage snippets using sub-pagesGoogle, 2013 · Active, term to 2033 · Inventors: Junli Xian, Harvey Jones, Samarth Keshava

  Rich snippets prefer Quality Nodes. Reference: Claim 1.

• US 10776435 B2 Canonicalized online document sitelink generationGoogle, 2017 · Active, term to 2037 · Inventors: Vaibhav Vaish, Venky Ramachandran, David Philip Sisson, Ramakrishnan Kandhan (+3 more)

  Sitelinks surface the site's Quality Nodes. Reference: Claim 1.

Google Content Warehouse leak 4 attributes

• QualityAuthorityTopicEmbeddingsVersionedItem.siteFocusScore Quality Nodes raise siteFocusScore by anchoring the topic.
• CompressedQualitySignals.authorityPromotion authority-boost flows to Quality Nodes linked from homepage.
• AnchorsAnchorSource.homePageInfo Quality Nodes linked from a FULLY_TRUSTED homepage inherit stronger authority.
• PerDocData.homepagePagerankNs homepage PR flows to Quality Nodes directly linked from it.

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Patent evidence 4 patents

• US 9165040 B1 Producing a ranking for pages using distances in a web-link graphGoogle, 2006 · Expired 2026 · Inventors: Nissan Hajaj

  Seed-distance ranking rewards clean Contextual Crawl Paths from Root hubs to Node pages. Reference: Claim 1.

• US 9953049 B1 Producing a ranking for pages using distances in a web-link graphGoogle, 2015 · Active, term to 2035 · Inventors: Nissan Hajaj

  Continuation.

• US 8825646 B1 / US 9400849 B1 Scalable system for determining short paths within web link networkGoogle, 2009 / 2014 · Active, term to 2029 · Inventors: Jesse Louis Alpert, Nissan Hajaj

  Compute shortest paths at scale, the infrastructure that reads crawl paths. Reference: Abstract.

• US 8121991 B1 Identifying transient paths within websitesGoogle, 2008 · Active, term to 2028 · Inventors: Eran Gabber, Michael Flaster, Ruoming Pang, Shanmugavelayutham Muthukrishnan

  Distinguishes persistent crawl paths from transient ones, Contextual Crawl Paths are persistent. Reference: Claim 1.

Google Content Warehouse leak 4 attributes

• IndexingDocjoinerDataVersion.indexingUrlPatternUrlTreeUrlPatternAnnotations URL-tree pattern annotations encode the crawl path hierarchy.
• IndexingSignalAggregatorUrlPatternSignals.coverage and .pagerankScore, URL-pattern signals measure how authoritatively the crawl path is used.
• IndexingMobileInterstitialsProtoDesktopInterstitials.urlTree explicit "url tree" concept at site-level.
• IndexingDocjoinerDataVersion.qualityBreadcrumbsBreadcrumbAnnotations breadcrumb annotations persist crawl paths.

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Patent evidence 8 patents

• US 7979417 B1 Embedded communication of link information (Reasonable Surfer)Google, 2005 · Expired 2025 · Inventors: Krishna Bharat, Matthew Daniel Cutts, Paul G. Haahr, Radhika A. Malpani (+2 more)

  Direct hit, this is the Reasonable Surfer / Cautious Surfer patent Koray cites. Background explicitly: "Web crawlers and related search engines are not provided with additional information that may be useful in determining a relative importance or weighting for one or more content locations", the patent describes embedded information that assigns weighting to each link. Reference: Background and claim 1.

• US 8260766 B2 Embedded communication of link informationGoogle, 2011 · Active, term to 2031 · Inventors: Krishna Bharat, Matthew Daniel Cutts, Paul G. Haahr, Radhika A. Malpani (+2 more)

  Granted continuation of US-7979417-B1 with the same link-weighting mechanism.

• US 11238211 B2 Automatic hyperlinking of documentsGoogle, 2019 · Active, term to 2039 · Inventors: Jan van de Kerkhof, Balint MIKLOS, Amr Abdelfattah, Tobias Kaufmann (+8 more)

  Link-likelihood table ("here" 0.3, "design doc" 0.8, "slides" 0.7). Reference: Link-likelihood table.

• US 9075778 B1 Predicting user navigation events within a browserGoogle, 2014 · Active, term to 2034 · Inventors: Dominic Hamon

  Per-link click probability, Reasonable Surfer weighting. Reference: Claim 1.

• US 8719276 B1 Ranking nodes in a linked database based on node independenceGoogle, 2011 · Active, term to 2031 · Inventors: Paul Haahr, Martin Kaszkiel, Amit Singhal

  Down-weights contributions from tightly interlinked (footer-link-farm) sources. Reference: Detailed description.

• US 7260573 B1 Personalizing anchor text scoresGoogle, 2004 · Expired 2024 · Inventors: Glen Jeh, Taher H. Haveliwala, Sepandar D. Kamvar

  Anchor-text scores from source documents contribute to target rank. Reference: Claim 1.

• US 8959093 B1 Ranking search results based on anchorsGoogle, 2010 · Active, term to 2030 · Inventors: Ajay P. Nerurkar

  Query-relevant anchor text propagates click evidence. Reference: Claim 1.

• US 8121991 B1 Identifying transient paths within websitesGoogle, 2008 · Active, term to 2028 · Inventors: Eran Gabber, Michael Flaster, Ruoming Pang, Shanmugavelayutham Muthukrishnan

  Persistent nav links carry more weight. Reference: Claim 1.

Google Content Warehouse leak 5 attributes

• AnchorsAnchor.pagerankWeight "Weight to be stored in linkmaps for pageranker." The actual link-level pagerank weight that sculpting manipulates.
• AnchorsAnchor.setiPagerankWeight alternate (temporary) PageRank weight.
• AnchorsAnchorSource.pagerank and .pagerankNs, source-page pagerank for the linking page.
• IndexingDocjoinerAnchorStatistics.topPrOnsiteAnchorCount "anchor with pagrank > 51000 is the best anchor." Direct Reasonable-Surfer-style tiering.
• CompositeDocForwardingDup.rawPagerank raw pagerank forwarding on duplicate URLs.

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Patent evidence 5 patents

• US 8700590 B2 Grouping multimedia and streaming media search resultsMicrosoft, 2012 · Active, term to 2032 · Inventors: Aram Christian Abajian

  Consolidates a plurality of URLs by matching URL character portions, the direct engineering of URL consolidation. Reference: Claim 1.

• US 8521746 B1 Detection of bounce pad sitesGoogle, 2011 · Active, term to 2031 · Inventors: Rupesh Kapoor, David Michael Proudfoot, Joachim Kupke

  Clusters duplicate documents, measures quality per document, ranks within the cluster. Direct patent on consolidation vs demotion within a duplicate cluster. Reference: Claim 1.

• US 7984054 B2 Representative document selection for sets of duplicate documentsGoogle, 2009 · Active, term to 2029 · Inventors: Daniel Dulitz, Alexandre A. Verstak, Sanjay Ghemawat, Jeffrey A. Dean

  Collapses duplicates into one representative. Reference: Claim 1.

• US 8280861 B1 Identifying duplicate electronic content based on metadataGoogle, 2011 · Active, term to 2031 · Inventors: Tim Park, Dmitry Dolinsky

  Hash-based duplicate identification. Reference: Claim 1.

• US 9081861 B2 URL canonicalizationGoogle, 2008 · Active, term to 2028 · Inventors: Matthew Lloyd, Thomas Bergan

  Direct URL canonicalization for consolidation. Reference: Claim 1.

Google Content Warehouse leak 4 attributes

• CompositeDocForwardingDup.purposes "The purpose(s) of the forwarding dup indicating if it is used for forwarding signal/anchors generally, or only for forwarding some specific signal (e.g. navboost), or for some other purposes." Explicit signal consolidation across duplicates.
• CompositeDocForwardingDup.rawPagerank raw PageRank forwarded via the dup forwarding.
• CompositeDocIndexingInfo.rawNavboost "The raw navboost count for the canonical url without aggregating the navboost from dup urls. This field is used when building forwarding map." Confirms canonical-level signal consolidation.
• PerDocData.urlAfterRedirectsFp fingerprint-based de-dup for redirect-chain consolidation.

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Patent evidence 5 patents

• US 9959315 B1 Context scoring adjustments for answer passagesGoogle, 2014 · Active, term to 2034 · Inventors: Nitin Gupta, Srinivasan Venkatachary, Lingkun Chu, Steven D. Baker

  Direct-answer passage scoring. Reference: Claim 1.

• US 11086959 B2 Secure digital assistant integration in web pagesGoogle, 2020 · Active, term to 2040 · Inventors: Adam Coimbra, Nikhil Rao, Hao Wu, Ian Baker (+2 more)

  Action JSON with intent fulfillment, directly implements Query Responsiveness for AI assistants. Reference: Example JSON.

• US 11829435 B2 Secure digital assistant integration in web pagesGoogle, 2022 · Active, term to 2042 · Inventors: Adam Coimbra, Nikhil Rao, Hao Wu, Ian Baker (+2 more)

  Continuation.

• US 9753901 B1 Identifying important document content using geometries of page elementsGoogle, 2013 · Active, term to 2033 · Inventors: Yifan Xu, Xiaofeng Mi

  Geometrically identifies the main interactive content. Reference: Abstract.

• US 10255380 B2 Identifying and structuring related dataMicrosoft, 2013 · Active, term to 2033 · Inventors: Jun Yan, Ning Liu, Lei Ji, Zheng Chen

  Related attributes displayed in a dedicated first UI region, Query Responsiveness in a structured container. Reference: Claim 1.

Google Content Warehouse leak 4 attributes

• QualityPreviewSnippetBrainFeatures.modelScore SnippetBrain evaluates how responsively the page answers the query.
• QualityPreviewSnippetDocumentFeatures.leadingtextDistanceScore direct responsiveness signal.
• IndexingDocjoinerDataVersion.indexingAnnotationsWaPassagesQueryToPassageAnnotations if a passage directly maps to a query, responsiveness is confirmed.
• IndexingDocjoinerDataVersion.indexingMlPageStructure ML page structure detects whether interactive components (calculators, forms) are present.

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Patent evidence 3 patents

• US 8407231 B2 Document scoring based on link-based criteriaGoogle, 2010 · Active, term to 2030 · Inventors: Anurag Acharya, Matt Cutts, Jeffrey Dean, Paul Haahr (+4 more)

  Link-appearance date profile; natural cadence rewarded. Reference: Claim 1.

• US 8549014 B2 Document scoring based on document content updateGoogle, 2011 · Active, term to 2031 · Inventors: Anurag Acharya, Jeffrey Dean, Paul Haahr, Monika Henzinger (+3 more)

  Score based on content update rate, natural cadence rewarded. Reference: Claim 1.

• US 9361369 B1 Clustering news online content based on content freshnessGoogle, 2012 · Active, term to 2032 · Inventors: Krishna Bharat, Michael Curtiss, Michael Schmitt

  Freshness scoring. Reference: Claim 1.

Google Content Warehouse leak 5 attributes

• IndexingDocjoinerDataVersion.firstseen first-seen date for crawl.
• AnchorsAnchor.firstseenDate firstseen date for discovered inlinks.
• QualityCopiaFireflySiteSignal.latestBylineDateSec site-level byline date.
• IndexingDocjoinerDataVersion.qualityNavboostCrapsCrapsPatternSignal "pattern signal" with potential to detect machine-publication rhythms.
• IndexingDocjoinerAnchorPhraseSpamInfo.phraseAnchorSpamRate anchor-rate spike detection; same pattern-detection family that catches artificial publication cadences.

Related research and notes

• Related: Scaled Content Abuse (Google Search Essentials docs). Not a patent; Google's own documentation on penalizing scaled, patterned publication.

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Patent evidence 1 patent

• US 7346839 B2 Information retrieval based on historical dataGoogle, 2003 · Expired 2023 · Inventors: Anurag Acharya, Matt Cutts, Jeffrey Dean, Paul Haahr (+6 more)

  Historical signals accrue to produce stable rankings. Reference: Claim 1 and claims 10-38.

Google Content Warehouse leak 4 attributes

• QualityNsrNsrData.siteQualityStddev "Estimate of site's PQ rating stddev." High stddev = unstable rankings = discontinuity.
• QualityNsrNsrData.chardVariance and .chardScoreVariance, "Site-level Chard Variance for all pages of a site." Variance measures discontinuity.
• CompressedQualitySignals.nsrConfidence "NSR confidence score." Low confidence = discontinuity.
• CompressedQualitySignals.nsrOverrideBid "NSR override bid, used in Q* for emergency overrides." Discontinuity triggers overrides.

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Patent evidence 5 patents

• US 8521746 B1 Detection of bounce pad sitesGoogle, 2011 · Active, term to 2031 · Inventors: Rupesh Kapoor, David Michael Proudfoot, Joachim Kupke

  Describes clusters of duplicate documents and quality-based consolidation, guardrail against fraudulent canonical swaps. Reference: Claim 1.

• US 8700590 B2 Grouping multimedia and streaming media search resultsMicrosoft, 2012 · Active, term to 2032 · Inventors: Aram Christian Abajian

  URL consolidation. Reference: Claim 1.

• US 9081861 B2 URL canonicalizationGoogle, 2008 · Active, term to 2028 · Inventors: Matthew Lloyd, Thomas Bergan

  Direct URL canonicalization. Reference: Claim 1.

• US 7984054 B2 Representative document selection for sets of duplicate documentsGoogle, 2009 · Active, term to 2029 · Inventors: Daniel Dulitz, Alexandre A. Verstak, Sanjay Ghemawat, Jeffrey A. Dean

  Representative selection, the receiving canonical in a migration. Reference: Claim 1.

• US 7346839 B2 Information retrieval based on historical dataGoogle, 2003 · Expired 2023 · Inventors: Anurag Acharya, Matt Cutts, Jeffrey Dean, Paul Haahr (+6 more)

  Historical trust is the asset being transferred. Reference: Claim 1.

Google Content Warehouse leak 3 attributes

• PerDocData.urlAfterRedirectsFp URL-after-redirects fingerprint, used to tie historical signals to canonicals through redirect chains.
• CompositeDocForwardingDup.purposes and .rawPagerank, explicit dup-forwarding mechanics behind cross-domain signal transfer.
• AnchorsAnchor.originalTargetDocid "The docid of the anchor's original target. This field is available if and only if the anchor is forwarded." Direct redirect-chain anchor preservation.

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Patent evidence 8 patents

• US 9026516 B1 Interest-based keywords for prominent entity and by locationGoogle, 2011 · Active, term to 2031 · Inventors: Charles Chapin, Vibhor Nanavati, Xuefu Wang, Xinyu Tang (+2 more)

  Targets local content based on user interests and current location, direct bridge between Search Action (query) and Real-World Action (visit). Reference: Detailed description.

• US 8312010 B1 Local business ranking using mapping informationGoogle, 2007 · Active, term to 2027 · Inventors: Johnny Chen, Clarence Chris Mysen

  Local business search; the primary Real-World-Action-oriented ranking patent. Reference: Background.

• US 7373246 B2 Using boundaries associated with a map view for business location searchingGoogle, 2005 · Expired 2025 · Inventors: Brian O'Clair

  Map-view boundary business retrieval, physical-world action target. Reference: Background.

• US 8068980 B2 Using boundaries for business location searchingGoogle, 2010 · Active, term to 2030 · Inventors: Brian O'Clair

  Continuation.

• US 11086959 B2 Secure digital assistant integration in web pagesGoogle, 2020 · Active, term to 2040 · Inventors: Adam Coimbra, Nikhil Rao, Hao Wu, Ian Baker (+2 more)

  ORDER_RIDE intent in the action JSON, Search Action triggering a Real-World Action (ride). Reference: Example JSON.

• US 11829435 B2 Secure digital assistant integration in web pagesGoogle, 2022 · Active, term to 2042 · Inventors: Adam Coimbra, Nikhil Rao, Hao Wu, Ian Baker (+2 more)

  Continuation.

• US 11960554 B2 Unsupervised entity and intent identificationMicrosoft, 2021 · Active, term to 2041 · Inventors: Gautam Prasad, Varun Appaswami, Bhanu Teja Chunduri

  Entity + intent extraction distinguishes transactional from informational intent, the Search vs Real-World Action split in query interpretation. Reference: Abstract.

• US 8423568 B2 Query classification using implicit labelsMicrosoft, 2009 · Active, term to 2029 · Inventors: Ariel D. Fuxman, Anitha Kannan, Andrew Brian Goldberg, Rakesh Agrawal

  Local-intent query identification. Reference: Detailed description.

Google Content Warehouse leak 4 attributes

• LocalWWWInfo.address .phone, .hours, real-world action signals (physical office intake).
• GeostoreEstablishmentProto.bizbuilderReference real-world claim of the business entity.
• QualityPreviewSnippetDocumentFeatures.leadingtextDistanceScore search-action responsiveness (answer at the top).
• IndexingDocjoinerDataVersion.indexingMlPageStructure detects whether calculators / tools (search-action) are present.

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Google Content Warehouse leak 4 attributes

• IndexingDocjoinerDataVersion.qualityTwiddlerDomainClassification "Twiddler Domain Classification." Direct per-category classifier.
• PerDocData.topPetacatTaxId "Top petacat of the site. Used in SiteboostTwiddler." Category-level quality signal.
• QualityNsrNsrData.chardEncoded .ymylNewsV2Score, .healthScore, .videoScore, .shoppingScore, per-vertical quality scores. These collectively comprise Categorical Quality.
• QualityNsrPQDataSubchunkData.type "Type of this chunk. Eg, ymyl_health, d2v, etc." Explicit category typing.

Related research and notes

• Content Warehouse API leak (May 2024): CategoricalQualityScore and CategoryParameters properties, leak-confirmed category-specific quality signal.

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Google Content Warehouse leak 3 attributes

• IndexingDupsLocalizedLocalizedCluster.boostSourceBlocker "a result which can be a boost target but should itself not be boosted." Composite-score management.
• CompositeDoc.docinfoPassthroughAttachments "used in post-doc twiddlers and for display in GWS." The twiddler layer is where composite SERP balance happens.
• PerDocData.datesInfo "Used in FreshnessTwiddler." Freshness twiddler contributes to composite SERP balancing.

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Patent evidence 2 patents

• US 8521746 B1 Detection of bounce pad sitesGoogle, 2011 · Active, term to 2031 · Inventors: Rupesh Kapoor, David Michael Proudfoot, Joachim Kupke

  Google's own patent on identifying bounce-pad sites that redirect to another destination, direct defense against malicious canonical hijacking. Reference: Claim 1.

• US 8886747 B1 Verifying domain ownershipGoogle, 2010 · Active, term to 2030 · Inventors: Bryan CHEN, Erik M. Kline, Marija Mikic-Rakic

  Verifies entity controls content for a domain name, the defensive primitive against hijack. Reference: Claim 1.

Google Content Warehouse leak 4 attributes

• PerDocData.urlAfterRedirectsFp redirect-chain fingerprint; defense against cross-domain canonical capture.
• CompositeDocAdditionalChecksums.SimHashIsTrusted "SimHash-is-trusted." Trust gate against spoofed duplicates.
• CompositeDocForwardingDup.purposes explicit purpose annotation on dup forwarding.
• AnchorsAnchor.originalTargetDocid preserves original target identity even after forwarding.

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Patent evidence 1 patent

• US 7346839 B2 Information retrieval based on historical dataGoogle, 2003 · Expired 2023 · Inventors: Anurag Acharya, Matt Cutts, Jeffrey Dean, Paul Haahr (+6 more)

  Historical data per host; subdomains accumulate separate histories. Reference: Claim 1.

Google Content Warehouse leak 4 attributes

• PerDocData.nsrSitechunk "SiteChunk computed for nsr. It some cases it can use more information than just url (e.g. youtube channels)." Hostname-level chunking confirms sub-domain partitioning works.
• QualityNsrNsrData.siteChunk + .siteChunkSource, sitechunk and source; sub-domain = separate sitechunk.
• QualityNsrNsrData.host "host" field used in NSR data; sub-domains get their own NSR row.
• QualityNsrNsrData.nsrdataFromFallbackPatternKey if a sub-domain lacks NSR data, it falls back to host average; partitioning fails if the subchunk cannot distinguish itself.

Related research and notes

• Content Warehouse API leak (May 2024): host-level hostAge and quality-signal fields in the leak are keyed by hostname, confirming subdomain-level scoring.

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Patent evidence 5 patents

• US 7984054 B2 Representative document selection for sets of duplicate documentsGoogle, 2009 · Active, term to 2029 · Inventors: Daniel Dulitz, Alexandre A. Verstak, Sanjay Ghemawat, Jeffrey A. Dean

  Representative selection, the receiving EMD. Reference: Claim 1.

• US 8521746 B1 Detection of bounce pad sitesGoogle, 2011 · Active, term to 2031 · Inventors: Rupesh Kapoor, David Michael Proudfoot, Joachim Kupke

  Clusters of equivalents, must not look like bounce pads during consolidation. Reference: Claim 1.

• US 7346839 B2 Information retrieval based on historical dataGoogle, 2003 · Expired 2023 · Inventors: Anurag Acharya, Matt Cutts, Jeffrey Dean, Paul Haahr (+6 more)

  Historical signals (age of document, age of links, anchor-text age) that chain through 301 redirects. Reference: Claim 1 and claims 10-38.

• US 8407231 B2 Document scoring based on link-based criteriaGoogle, 2010 · Active, term to 2030 · Inventors: Anurag Acharya, Matt Cutts, Jeffrey Dean, Paul Haahr (+4 more)

  Appearance date per link, preserved through 301 redirects. Reference: Claim 1.

• US 9081861 B2 URL canonicalizationGoogle, 2008 · Active, term to 2028 · Inventors: Matthew Lloyd, Thomas Bergan

  Canonicalizes URL parameters, a signal-chain prerequisite. Reference: Claim 1.

Google Content Warehouse leak 4 attributes

• PerDocData.hostAge host-age signal; aged domains carry high hostAge.
• IndexingDocjoinerDataVersion.spamContentSiteAgeStats site-age statistics inspected during spam filtering.
• CompositeDocForwardingDup.rawPagerank PageRank forwarded through 301 chains.
• AnchorsAnchor.firstseenDate anchors inherit firstseen dates through chaining.

Related research and notes

• Note: Google deprecated AMP's SERP preference in 2022; AMP-era historical signals are preserved through canonical/redirect but the preferential surface no longer exists. Chaining still works on the signal-preservation axis Koray describes.

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Patent evidence 5 patents

• US 11361227 B2 Onboarding of entity dataGoogle, 2017 · Active, term to 2037 · Inventors: Ion Constantinescu

  Direct hit. Claim 1: computing system associated with an existing knowledge graph receives request to onboard entity data. The patent that defines how a Web Entity becomes a Knowledge Graph entity. Reference: Claim 1 and claim 7 (attribute-to-entity associations).

• US 11769064 B2 Onboarding of entity dataGoogle, 2022 · Active, term to 2042 · Inventors: Ion Constantinescu

  Continuation.

• US 12001799 B1 Optimized graph traversalGoogle, 2023 · Active, term to 2043 · Inventors: Christopher Durr, Hector Mauricio Ayala, Mayank Jain

  Entity's properties accessed by graph traversal of its connected nodes, Web Entities surface their attributes through this traversal. Reference: Claim 13.

• US 9665643 B2 Knowledge-based entity detection and disambiguationMicrosoft, 2011 · Active, term to 2031 · Inventors: Kang Li, Yi Li, Yiping Zhou, Zhengdong Lv (+1 more)

  Tags named entities on web pages and uses that for organization. Reference: Detailed description.

• US 8886747 B1 Verifying domain ownershipGoogle, 2010 · Active, term to 2030 · Inventors: Bryan CHEN, Erik M. Kline, Marija Mikic-Rakic

  Verifies entity controls content for a domain, ties Web Entity to a verifiable owner. Reference: Claim 1.

Google Content Warehouse leak 4 attributes

• GeostoreKnowledgeGraphReferenceProto.id "KG Identifier (MID)." The entity's primary Knowledge-Graph ID.
• GeostoreEstablishmentProto.bizbuilderReference "indicates that a feature is claimed in CBDB (with the canonical state in MapFacts)." Verified Web Entity status for local businesses.
• IndexingDocjoinerDataVersion.repositoryWebrefSalientEntitySet Web Entity's salient-entity footprint per page.
• RepositoryWebrefEntityAnnotations.detailedEntityScores "Additional information about how the entity relates to the page, for example whether it is a business entity which published the page." Direct "is this the brand's own page" annotation.

Related research and notes

• Wikipedia / Wikidata / Freebase (foundational KG datasets Google and Microsoft built atop).

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Patent evidence 5 patents

• US 10394816 B2 Detecting product lines within product search queriesGoogle, 2012 · Active, term to 2032 · Inventors: Ritendra Datta

  Pairs detected product lines with product brands, brand-query recognition pipeline. Reference: Technical field.

• US 10354292 B1 / US 9378276 B1 Systems and methods for generating navigation filtersGoogle, 2016 / 2014 · Active, term to 2036 · Inventors: Shreyas Doshi, Venky Ramachandran, Gaurav Garg, Poorva Hari Potdar (+3 more)

  Navigation filters derived from brand-anchored headings. Reference: Detailed description brands_by_entities config.

• US 8959093 B1 Ranking search results based on anchorsGoogle, 2010 · Active, term to 2030 · Inventors: Ajay P. Nerurkar

  Query-specific click behavior propagated between pages. Reference: Claim 1.

• US 7346839 B2 Information retrieval based on historical dataGoogle, 2003 · Expired 2023 · Inventors: Anurag Acharya, Matt Cutts, Jeffrey Dean, Paul Haahr (+6 more)

  Historical click-selection data. Reference: Claim 1 and claims 10-38.

• US 8886747 B1 Verifying domain ownershipGoogle, 2010 · Active, term to 2030 · Inventors: Bryan CHEN, Erik M. Kline, Marija Mikic-Rakic

  Verifies the brand owner, the identity anchor Navboost associates with click evidence. Reference: Claim 1.

Google Content Warehouse leak 6 attributes

• QualityNavboostCrapsCrapsClickSignals.clicks .goodClicks, .lastLongestClicks, raw brand+topic click metrics.
• QualityNavboostCrapsCrapsClickSignals.absoluteImpressions impressions for brand queries.
• QualityNavboostCrapsCrapsData.country brand-search demand is country-sliced.
• QualityNsrNsrData.pnavClicks "denominator for the pnav computation." pnav aggregates navigational clicks sitewide.
• QualityNsrNsrData.pnav fractional navigational signals (brand nav).
• QualityNsrNsrData.chromeInTotal Chrome direct-navigation total; direct-brand-type-in traffic.

Related research and notes

• Content Warehouse API leak (May 2024): NavBoost module explicitly named, with goodClicks, badClicks, lastLongestClicks, unsquashedClicks, squashedClicks properties, leak-confirmed Brand Search Demand mechanism.
• Reference: US v. Google DOJ antitrust testimony (Pandu Nayak, 2023), publicly disclosed existence of Navboost. Not a patent; evidentiary.

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Patent evidence 3 patents

• US 8682913 B1 Corroborating facts extracted from multiple sourcesGoogle, 2005 · Expired 2025 · Inventors: Jonathan T. Betz

  Direct hit. Abstract: "A method and system for corroborating facts extracted as attribute-value pairs from multiple sources. Facts associated with a common subject are identified and extracted from multiple sources. From the extracted facts, attributes satisfying a first [consensus criterion]." The exact Andrew Hogue Browsable Fact Repository patent Koray cites. Reference: Abstract, Claim 1 (identify source documents, identify attribute-value pairs, corroborate).

• US 8825471 B2 Unsupervised extraction of factsGoogle, 2006 · Expired 2026 · Inventors: Jonathan T. Betz, Shubin Zhao

  Seed attribute-value pair extracted from one document, then verified across second documents. Reference: Abstract.

• US 10354188 B2 / US 11475319 B2 Extracting facts from unstructured informationMicrosoft, 2016 / 2019 · Active, term to 2036 · Inventors: Achraf Abdel Moneim Tawfik Chalabi, Ahmed Mohamed Emad Morsi Abdelbaki, Brandon Robert Anderson, Eslam Kamal Abdel-Aal Abdel-Reheem (+4 more)

  Fact extraction pipeline that feeds consensus. Reference: Abstract.

Google Content Warehouse leak 4 attributes

• IndexingDocjoinerDataVersion.repositoryAnnotationsSporeAuthoritySignal Spore authority signal reconciles facts across sources.
• IndexingDocjoinerDataVersion.repositoryProtiumGoldmineTriplificationResult triples from each source are cross-referenced.
• IndexingDocjoinerDataVersion.researchStructuredDataKnowledgeExplorationAnnotations knowledge-exploration annotations consolidate structured data.
• RepositoryWebrefAggregatedEntityNameScores.entityScore entity-name scores aggregated across the web.

Related research and notes

• Knowledge Vault (Dong et al., KDD 2014), probabilistic multi-source fact fusion.
• Knowledge-Based Trust (Dong et al., VLDB 2015, arXiv:1502.03519). Explicit corroboration-based trust scoring.
• Attributed Question Answering (Bohnet et al., arXiv:2212.08037).

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Patent evidence 2 patents

• US 8682913 B1 Corroborating facts extracted from multiple sourcesGoogle, 2005 · Expired 2025 · Inventors: Jonathan T. Betz

  Factual-corroboration mechanism. KBT is corroboration + confidence propagated to source. Reference: Abstract.

• US 11361227 B2 Onboarding of entity dataGoogle, 2017 · Active, term to 2037 · Inventors: Ion Constantinescu

  Trust gating when a new entity submission is checked against the existing graph. Reference: Claim 1.

Google Content Warehouse leak 4 attributes

• IndexingDocjoinerDataVersion.repositoryAnnotationsSporeAuthoritySignal direct "authority signal" from the structured-data pipeline that feeds KG; closest leak analog to KBT output.
• IndexingDocjoinerDataVersion.researchStructuredDataKnowledgeExplorationAnnotations structured-data KBT annotation.
• IndexingDocjoinerDataVersion.repositoryProtiumGoldmineTypeAssertionOutput type assertions that match KG raise KBT.
• GeostoreTrustSignalsProto.sourceTrust source-trust signal.

Related research and notes

• KELM (Agarwal et al., arXiv:2010.12688, 2020). Knowledge-Enhanced Language Model, direct consumer of KBT-grade facts.

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Patent evidence 8 patents

• US 10599770 B1 Generating author vectorsGoogle, 2018 · Active, term to 2038 · Inventors: Brian Patrick Strope, Quoc V. Le

  Direct hit. Claim 1: "receiving an input from a first author; obtaining an author vector for the first author, the author vector characterizing at least one of a communication style, a personality, [or] of the first author." The exact "author vector" Koray cites. Reference: Claim 1.

• US 9984062 B1 Generating author vectorsGoogle, 2016 · Active, term to 2036 · Inventors: Brian Patrick Strope, Quoc V. Le

  Earlier grant. Trains per-author embeddings from sequences of words. Reference: Claim 1.

• US 11868724 B2 Generating author vectorsGoogle, 2022 · Active, term to 2042 · Inventors: Quoc V. Le, Brian Patrick Strope

  Continuation; uses author-identifier input to a machine-learned model. Reference: Claim 1.

• US 9400789 B2 Associating resources with entities (Author Profiles)Google, 2013 · Active, term to 2033 · Inventors: James T. Connor, Alexandre A. Verstak, Anurag Acharya

  Claim 1: identifies information to add to an author profile, classifies documents as authored by that author, assigns documents to document groups. Direct patent on author-profile construction. Reference: Claim 1.

• US 11361227 B2 Onboarding of entity dataGoogle, 2017 · Active, term to 2037 · Inventors: Ion Constantinescu

  Onboarding of entity data into the KG, Entity Home is the owner-asserted entry point. Reference: Claim 1.

• US 8886747 B1 Verifying domain ownershipGoogle, 2010 · Active, term to 2030 · Inventors: Bryan CHEN, Erik M. Kline, Marija Mikic-Rakic

  Domain-ownership verification underpins Entity Home. Reference: Claim 1.

• US 12001799 B1 Optimized graph traversalGoogle, 2023 · Active, term to 2043 · Inventors: Christopher Durr, Hector Mauricio Ayala, Mayank Jain

  Entity properties resolved by graph traversal from an Entity Home node. Reference: Claim 13.

• US 9665643 B2 Knowledge-based entity detection and disambiguationMicrosoft, 2011 · Active, term to 2031 · Inventors: Kang Li, Yi Li, Yiping Zhou, Zhengdong Lv (+1 more)

  Disambiguates named entities, Entity Home provides the canonical disambiguation anchor. Reference: Detailed description.

Google Content Warehouse leak 6 attributes

• Author Vectors
• IndexingDocjoinerDataVersion.qualityAuthorshipAuthorAnnotations explicit per-document authorship annotation.
• NlpSaftDocument.bylineDate "Document's byline date, if available."
• QualityCopiaFireflySiteSignal.latestBylineDateSec site-level byline consistency.
• IndexingDocjoinerDataVersion.qualityDniDocPreviewRestrictions → carries bylineDateSecs, explicit webmaster-declared author/publish date.
• Entity Homes
• RepositoryWebrefEntityAnnotations.detailedEntityScores includes "whether it is a business entity which published the page." The Entity Home is the page where this flag is true.
• GeostoreKnowledgeGraphReferenceProto.id the MID bound to the Entity Home.

Related research and notes

• Reference: Jason Barnard, "Entity Home" doctrine (2019-2024 Kalicube publications). Not a patent; original articulation Koray extends.

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Patent evidence 3 patents

• US 11983173 B2 Schema-agnostic query templatesGoogle, 2019 · Active, term to 2039 · Inventors: Anastasios Kementsietsidis, Jay Yogeshbhai Pandya, Tingting TANG, Laurren Kanner

  Query templates feeding Document Ranking. Reference: Abstract.

• US 9959315 B1 Context scoring adjustments for answer passagesGoogle, 2014 · Active, term to 2034 · Inventors: Nitin Gupta, Srinivasan Venkatachary, Lingkun Chu, Steven D. Baker

  Passage Ranking patent. Reference: Claim 1.

• US 11409748 B1 Context scoring adjustments for answer passagesGoogle, 2018 · Active, term to 2038 · Inventors: Nitin Gupta, Srinivasan Venkatachary, Lingkun Chu, Steven D. Baker

  Continuation.

Google Content Warehouse leak 3 attributes

• Document ranking → CompressedQualitySignals.siteAuthority, .lowQuality, .babyPandaV2Demotion, QualityNsrNsrData.nsr.
• Passage ranking → IndexingDocjoinerDataVersion.indexingAnnotationsWaPassagesQueryToPassageAnnotations.
• Passage generation → QualityPreviewSnippetBrainFeatures.modelScore + QualityPreviewSnippetDocumentFeatures.* (the "preview" family generates passage snippets).

Related research and notes

• REALM (Guu et al., ICML 2020). Document + passage retrieval before generation, the SERP-Triad research foundation.
• Attributed Question Answering (Bohnet et al., arXiv:2212.08037). Grounded generation with citations.
• FreshLLMs (Vu et al., arXiv:2310.03214). Retrieval + generation freshness.
• Reference: Google AI Overviews, Gemini, Perplexity, ChatGPT Search, product-side implementations of the SERP Triad Koray references.

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Patent evidence 3 patents

• US 8682913 B1 Corroborating facts extracted from multiple sourcesGoogle, 2005 · Expired 2025 · Inventors: Jonathan T. Betz

  The source corroboration mechanism that reads Surround Sound signals. Reference: Abstract.

• US 11361227 B2 Onboarding of entity dataGoogle, 2017 · Active, term to 2037 · Inventors: Ion Constantinescu

  Entity attribute-to-entity binding uses co-occurrence as evidence. Reference: Claim 7.

• US 9400789 B2 Associating resources with entities (Author Profiles)Google, 2013 · Active, term to 2033 · Inventors: James T. Connor, Alexandre A. Verstak, Anurag Acharya

  Author-to-document association, Surround Sound at the author level. Reference: Claim 1.

Google Content Warehouse leak 4 attributes

• IndexingDocjoinerDataVersion.repositoryAnnotationsSporeAuthoritySignal co-occurrence authority sampled from the entire web ecosystem.
• RepositoryWebrefAggregatedEntityNameScores.entityScore aggregated entity-name score across the web (surround-sound evidence).
• AnchorsAnchor.encodedNewsAnchorData "Encoded data containing information about newsiness of anchor. Populated only if anchor is classified as coming from a newsy, high quality site." Surround-sound citations from news sources get special treatment.
• IndexingDocjoinerAnchorTrustedInfo.trustedScore "Fraction of pages with newsy anchors on the site, >0 for trusted sites."

Related research and notes

• Barnacle SEO / Surround Sound SEO, Jason Barnard, 2018-2024 Kalicube publications. The marketing-side articulation.

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Patent evidence 2 patents

• US 8868402 B2 Construction of text classifiersGoogle, 2009 · Active, term to 2029 · Inventors: Dmitry Korolev, Hartmut Maennel, Matthias Heiler, Michael Schaer (+3 more)

  "Web site scores used to generate candidate phrases from a collection of Web documents", the site-level term vector construction. Reference: Detailed description.

• US 7177796 B1 Automated set up of web-based natural language interfaceMicrosoft, 2000 · Expired 2020 · Inventors: Frederick J. Damerau, David E. Johnson

  Automatically induces a topic hierarchy by examining web site structure and creates n-grams per topic, per-site term vector. Reference: Claim 1.

Google Content Warehouse leak 4 attributes

• QualityNsrNsrData.site2vecEmbeddingEncoded site2vec is the vector representation that encodes the Source Term Vector.
• QualityAuthorityTopicEmbeddingsVersionedItem.siteEmbedding site-level embedding.
• QualityAuthorityTopicEmbeddingsVersionedItem.siteFocusScore how focused the vector is around a single topic.
• PerDocData.topPetacatTaxId site's top category; crudely a vector label.

Related research and notes

• Content Warehouse API leak (May 2024): siteEmbeddings, pageEmbeddings, siteFocusScore, siteRadius, the leak-confirmed Source Term Vector properties.
• Jeffrey Dean indexing papers (Dean & Barroso IEEE Micro 2003; Jouppi/Dean TPU papers), foundational Google infrastructure referenced by Koray.

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Patent evidence 3 patents

• US 11074293 B2 Generating probabilistic transition dataMicrosoft, 2014 · Active, term to 2034 · Inventors: Sushma Bannur, Omar Alonso

  Direct patent on probabilistic transition graphs for user transitions. Reference: Abstract.

• US 10387437 B2 Query rewriting using session informationGoogle, 2017 · Active, term to 2037 · Inventors: Marcin M. Nowak-Przygodzki, Behshad Behzadi

  Session-informed query rewriting, Markovian treatment of query sessions. Reference: Claim 1.

• US 9547690 B2 Query rewriting using session informationGoogle, 2014 · Active, term to 2034 · Inventors: Marcin M. Nowak-Przygodzki, Behshad Behzadi

  Earlier grant.

Google Content Warehouse leak 4 attributes

• QualityNavboostCrapsCrapsClickSignals.lastLongestClicks "last longest clicks" captures end-state of a search session (satisfactory final click).
• QualityNavboostCrapsCrapsClickSignals.goodClicks and .badClicks, click-quality signals that classify state transitions.
• IndexingDocjoinerDataVersion.qualityNavboostCrapsCrapsPatternSignal pattern signal across session states.
• IndexingSignalAggregatorAgeWeightedCoverageData.luDwells LocalUniversal dwell counts; dwell-level state tracking.

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Patent evidence 1 patent

• US 8930822 B2 Method for human-centric information access and presentationMicrosoft, 2009 · Active, term to 2029 · Inventors: Øystein Haug Olsen

  Discusses ranking navigators based on user data and information content, direct application of Information Foraging. Reference: Conclusions section.

Google Content Warehouse leak 3 attributes

• QualityPreviewSnippetDocumentFeatures.leadingtextDistanceScore distance to information "prey."
• QualityNavboostCrapsCrapsClickSignals.lastLongestClicks foraging-success signal.
• QualityPreviewSnippetBrainFeatures.modelScore SnippetBrain predicts foraging reward.

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Patent evidence 1 patent

• US 8898156 B2 Query expansion for web searchMicrosoft, 2011 · Active, term to 2031 · Inventors: Jun Xu, Hang Li

  Re-ranking model operating in query-space and document-space, different queries produce different document rankings. Reference: Fig. 2 (query space vs document space).

Google Content Warehouse leak 2 attributes

• CompressedQualitySignals.experimentalQstarSignal .experimentalNsrTeamData, .experimentalNsrTeamWsjData, multiple experimental scoring overrides; different engines consume different signals.
• QualityAuthorityTopicEmbeddingsVersionedItem.versionId multiple embedding versions coexist; different consumers may read different versions.

Related research and notes

• FreshLLMs (Vu et al., arXiv:2310.03214, 2023). Different LLMs behave differently on recency. Direct research grounding for Algorithmic Relativism.
• REALM vs DPR vs ColBERTv2, different retrieval models yield different top-k. Multiple architectures produce different results on the same corpus.

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Patent evidence 3 patents

• US 8868402 B2 Construction of text classifiersGoogle, 2009 · Active, term to 2029 · Inventors: Dmitry Korolev, Hartmut Maennel, Matthias Heiler, Michael Schaer (+3 more)

  Site-wide n-gram features for topical classification, direct Source-Level Feature Vector mechanism. Reference: Detailed description.

• US 7177796 B1 Automated set up of web-based natural language interfaceMicrosoft, 2000 · Expired 2020 · Inventors: Frederick J. Damerau, David E. Johnson

  Site-level topical hierarchy + n-gram features. Reference: Claim 1.

• US 8407231 B2 Document scoring based on link-based criteriaGoogle, 2010 · Active, term to 2030 · Inventors: Anurag Acharya, Matt Cutts, Jeffrey Dean, Paul Haahr (+4 more)

  Site-wide link-based features. Reference: Claim 1.

Google Content Warehouse leak 4 attributes

• CompressedQualitySignals.siteAuthority the compressed site-level feature.
• QualityNsrNsrData whole module, 40+ site-level features: nsr, localityScore, tofu, chardEncoded, videoScore, shoppingScore, healthScore, pnav, impressions, ymylNewsV2Score, siteAutopilotScore, siteQualityStddev, smallPersonalSite, chromeInTotal, racterScores, siteLinkOut, vlq, clutterScores, articleScoreV2.
• QualityAuthorityTopicEmbeddingsVersionedItem site/page embeddings and site focus score.
• IndexingDocjoinerDataVersion.qualityQ2SiteQualityFeatures the master Q2 site-quality feature aggregation.

Related research and notes

• Content Warehouse API leak (May 2024): siteAuthority, siteFocusScore, siteRadius, siteEmbeddings, pageEmbeddings, chromeInTotal, ymylHealthScore, ymylNewsScore, leak-confirmed Source-Level Feature Vector properties.

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Patent evidence 2 patents

• US 8682913 B1 Corroborating facts extracted from multiple sourcesGoogle, 2005 · Expired 2025 · Inventors: Jonathan T. Betz

  The consumer of External Topical Map data, fact extraction from third-party sources. Reference: Abstract.

• US 11361227 B2 Onboarding of entity dataGoogle, 2017 · Active, term to 2037 · Inventors: Ion Constantinescu

  Third-party entity data onboarded into KG, directly rewards External Topical Maps. Reference: Claim 1.

Google Content Warehouse leak 4 attributes

• IndexingDocjoinerDataVersion.repositoryAnnotationsSporeAuthoritySignal authority accumulates from external sources via Spore.
• AnchorsAnchor.encodedNewsAnchorData news-anchor data from third-party sites.
• IndexingDocjoinerAnchorTrustedInfo.trustedScore site-wide trusted-source score built from external citations.
• RepositoryWebrefAggregatedEntityNameScores.entityScore web-wide entity-name score across third-party platforms.

Related research and notes

• Barnacle SEO / Surround Sound SEO (Jason Barnard Kalicube, 2018+). The marketing-side articulation.

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Patent evidence 3 patents

• US 8682913 B1 Corroborating facts extracted from multiple sourcesGoogle, 2005 · Expired 2025 · Inventors: Jonathan T. Betz

  Direct patent-side mechanism for consensus-based fact verification across sources, the engine Consensus Building targets. Reference: Abstract, Claim 1.

• US 9594831 B2 Targeted disambiguation of named entitiesMicrosoft, 2012 · Active, term to 2032 · Inventors: Chi Wang, Kaushik Chakrabarti, Tao Cheng, Surajit Chaudhuri

  Identifies named-entity occurrences across documents, unlinked brand mentions enter the consensus. Reference: Background.

• US 11361227 B2 Onboarding of entity dataGoogle, 2017 · Active, term to 2037 · Inventors: Ion Constantinescu

  Attribute-to-entity binding with multiple-source weighting. Reference: Claim 7.

Google Content Warehouse leak 4 attributes

• IndexingDocjoinerDataVersion.repositoryAnnotationsSporeAuthoritySignal Spore extracts and reconciles the same factual triple from unlinked sources, which is pure consensus.
• RepositoryWebrefAggregatedEntityNameScores.entityScore mention-frequency score across the web (no link required).
• IndexingDocjoinerDataVersion.repositoryProtiumGoldmineTriplificationResult triples from unlinked mentions still enter the KG.
• AnchorsAnchor.encodedNewsAnchorData even without passing PageRank, newsy mentions carry weight.

Related research and notes

• Knowledge-Based Trust (Dong et al., VLDB 2015, arXiv:1502.03519). The explicit "consensus not links" research foundation.
• Knowledge Vault (Dong et al., KDD 2014). Fact fusion from many sources without regard to hyperlinks.
• Luna Dong Stanford lecture on KELM, Koray's cited source for Consensus Building theory.

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Patent evidence 4 patents

• US 8458196 B1 System and method for determining topic authorityGoogle, 2012 · Active, term to 2032 · Inventors: Michael Jeffrey Procopio

  Direct hit. Claim 1: "determining topical authority, comprising: receiving… topic information for a document, the information including at least one topic and a weight for each topic." Google's own "Topic Authority" patent, the engineering substrate Topical Ownership builds on. Reference: Claim 1.

• US 11657223 B2 Keyphrase extraction beyond language modelingMicrosoft, 2021 · Active, term to 2041 · Inventors: Li Xiong, Chuan Hu, Arnold OVERWIJK, Junaid Ahmed (+2 more)

  Detailed description: "to produce topical authority ranking at web or internet scale… ordering and presenting [results]." Direct patent naming "topical authority ranking." Reference: Detailed description, Figs. 2A/2B.

• US 11361227 B2 Onboarding of entity dataGoogle, 2017 · Active, term to 2037 · Inventors: Ion Constantinescu

  Topical Ownership = owning the entity's canonical attributes in the KG. Reference: Claim 1 and Claim 7.

• US 9158849 B2 Synthesis of webpage snippetsGoogle, 2013 · Active, term to 2033 · Inventors: Junli Xian, Harvey Jones, Samarth Keshava

  Rich-snippet surfacing from one dominant source. Reference: Claim 1.

Google Content Warehouse leak 4 attributes

• CompressedQualitySignals.authorityPromotion the "promotion" bit for sources that own their topic.
• PerDocData.nsrIsCovidLocalAuthority .nsrIsElectionAuthority, IndexingDocjoinerDataVersion.localsearchAuthoritySiteAnnotation, explicit authority-site flags = Topical Ownership codified for whitelisted verticals.
• QualityNsrNsrData.siteAuthority site-level authority = Topical Ownership score.
• QualityAuthorityTopicEmbeddingsVersionedItem.siteFocusScore ownership demands high topic focus.

Related research and notes

• "Why topical authority isn't enough for AI search", Jason Barnard & Koray, Search Engine Land Aug 2026. The marketing-side articulation.

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.

Patent evidence 2 patents

• US 8682913 B1 Corroborating facts extracted from multiple sourcesGoogle, 2005 · Expired 2025 · Inventors: Jonathan T. Betz

  Wasteful Domains seed facts into this corroboration engine. Reference: Abstract.

• US 11361227 B2 Onboarding of entity dataGoogle, 2017 · Active, term to 2037 · Inventors: Ion Constantinescu

  Wasteful Domains supply attribute triples that onboard into the KG. Reference: Claim 1.

Google Content Warehouse leak 5 attributes

• CompositeDocForwardingDup.purposes forwarding-dup purpose annotation (signal routing).
• AnchorsAnchor.sourceType anchors from Extension Domains contribute per-tier.
• AnchorsAnchorSource.homePageInfo homepage-trust tier of the Extension Domain affects the equity it transfers.
• IndexingDocjoinerDataVersion.repositoryAnnotationsSporeAuthoritySignal a Wasteful Domain's job is to seed triples into Spore for corroboration.
• IndexingDocjoinerDataVersion.repositoryProtiumGoldmineTriplificationResult Wasteful Domains feed Goldmine with corroborating triples.

Patents and leaked attributes document mechanisms Google and Microsoft built. They are evidence of how the machinery is designed, not a guarantee of the exact live ranking weights today.