Every fact arrives with its passport and goes through arbitration before it enters the Knowledge Graph
Google gives the order of magnitude itself: 5 billion entities and 500 billion facts. The ruling in the US antitrust case describes data coming “from over [redacted] data feeds and pipelines, including from third parties”, processed “with the assistance of thousands of employees and contractors”. Google’s internal schemas show how those facts get in, how Google chooses between two sources that disagree, and what it keeps when it corrects or deletes.
Google’s AI uses it: AI Overviews work “in tandem” with the graph and AI Mode taps into “facts from our Knowledge Graph”, according to a Google document from May 2025. The Gemini app receives parts of it “that are not available to third parties”, the court ruling says. For a brand, being a well-described entity in the graph therefore weighs in Gemini’s and AI Mode’s answers.
products may use a fact
partners, editors, the web
a fact, when the feed requires it
The five Google systems involved, in the order they act
- Collection, from the webKnowledge Vault
What it does Extracts facts from web pages (text, tables, markup) and gives each one a probability of being true. SAFT and Tractzor are two of its extractors.
Its goal Fill the graph from the web.
- Collection, by peopleHume
What it does The graph’s editing tool, where Google employees and contractors fix and complete facts.
Its goal Fix and complete the graph by hand.
- WritingLivegraph
What it does Receives each fact and attaches it to an entity. Its Composer chooses between two conflicting values, its FactStore holds the facts and infers new ones.
Its goal Write facts into the graph, tied to the right entity and settled.
- ServingTopicServer
What it does Delivers entities and their facts to Google’s products: knowledge panel, local pack, Gemini…
Its goal Give each product the facts it is allowed to use.
- DisplayRefX
What it does Decides, for a query, whether to show an entity result such as a knowledge panel, relying on the graph’s facts.
Its goal Trigger the entity box that answers the query.
Our finding
Every Knowledge Graph fact carries a passport that names its source
A Knowledge Graph fact has three parts: a subject, a predicate, an object. “Notre-Dame de Paris fire, start date, 15 April 2019” is one. Google always stores that triple with its provenance, which we call its passport here: who wrote it, where it comes from, under which contract, and what Google’s systems may do with it. The passport first names the feed, the channel the fact came through: a data supplier under contract, a program that extracts facts from web pages, a team of editors, or another Google database such as the Maps one.
Knowledge Graph · a fact’s passport
- Subject
- Notre-Dame de Paris fire
/g/11h10vl01n - Predicate
- start date
- Object
- 15 April 2019
Visas: the fields of its provenance
Pick a stamp.
source_categoryThe source family: THIRD_PARTY (an outside supplier), CURATION (human editors), PARTNER_FEED (a partner’s feed) or EXTRACTION (a program that extracts facts, mostly from web pages).
Engineers’ commentUsed to measure impact of 3P contributions.
processThe name of the feed that wrote the fact: “amg” and “tms” for All Music Guide and Tribune Media Services, long-standing suppliers of music and TV data, SAFT or Tractzor for two extractors of Knowledge Vault.
Engineers’ commentUsed to uniquely identify data sources. Freebase: the OAuth application KG: the name of the source directory the triple was loaded from (eg, "amg", "tms ", "collections"). KV: identifier of an extraction system, e.g., SAFT or Tractzor.
creatorThe feed’s owner: the internal ID of the engineer or team that set up the import. Every feed has a named owner.
Engineers’ commentKG: the Google LDAP of the developer or MDB group that set up the triplification and import pipeline for this data source.
sourceThe URL of the page where the fact was found. Its host, that is the site name (www.example.com), is used to count sources when a feed requires triangulation: confirmation by three distinct sites.
Engineers’ commentIf the triples were extracted from the web, the source URL where the assertion was found. TopicServer will serve/populate/retain if REQUIRES_CITATION is also set, or if the process is explicitly allowed.
access_requiredThe access requirement, Google’s term: the contract or legal rule that governs the fact. Only the Google products that meet this requirement see the fact. 194 possible requirements, listed just below.
Engineers’ commentSpecifies the contract or legal visibility required to see the Triple.
is_supporting_dataA flag set on clues used only to recognise the entity, such as a phone number or an address: Google never displays them.
Engineers’ commentIndicates that the corresponding data is supporting evidence for reconciliation only, and is *not* an assertion that should be visible to other systems or to external users.
requires_triangulationA flag that requires three distinct sources before the fact is served. A source is counted by its site name: three pages on one site count as one.
Engineers’ commentThis indicates that the triple needs to be triangulated by 3 different sources before it can be served in production. A triple’s source is identified by its URL(provenance.source).host(). […] If you’re a new client trying to enable triangulation for your feed, please contact […] before using this feature.
weak_dataA flag that puts the data behind the rest: a new feed enters the graph without displacing the data already in place.
Engineers’ commentThis allows ingesting data with a lower chance of negatively affecting existing features and products, at the cost of potentially not showing the data when competing data is available.
deletion_provenanceA deletion carries a passport too: the graph dates every removal (deleted_on) and records who asked for it.
access requirements say which Google products may use a fact, and how far they go with it. The other seven values are technical or retired.
The 194 access requirements list the graph’s suppliers and the data Google keeps apart
An access requirement says which Google products may use a fact. Netflix’s catalogue arrives with the CONTRACTUAL_NETFLIX_DATA requirement: only the products covered by the Netflix contract see it. The Content Warehouse leak, Google’s internal documentation published by mistake in 2024, only gave the field’s name, access_required. The schemas give its 194 values, and with them the list of the graph’s suppliers and of the uses it restricts.
Who sees what: the products that access requirements clear
A requirement’s name often says which Google product the fact may be used in, or where it must never show.
product clearedproduct excluded
| Access requirement |  Search |  Assistant |  Google TV, Android TV |  YouTube |  Google Play |  Maps |  Shopping |  Ads |
|---|---|---|---|---|---|---|---|---|
| Netflix | ||||||||
the catalogue, in SearchCONTRACTUAL_ | cleared | |||||||
the catalogue, on Android TVCONTRACTUAL_ | cleared | |||||||
the “watch on Netflix” buttonNETFLIX_ | cleared | cleared | ||||||
| Media and service contracts | ||||||||
the Assistant’s media actionsCONTRACTUAL_ | cleared | |||||||
food ordering through the AssistantCONTRACTUAL_ | cleared | |||||||
Google TV’s free channelsCONTRACTUAL_ | cleared | |||||||
Google Play filmsCONTRACTUAL_ | cleared | |||||||
YouTube shows before releaseCONTRACTUAL_ | cleared | |||||||
| Shopping | ||||||||
free listings, never adsSHOPPING_ | cleared | excluded | ||||||
only on google.comSHOPPING_ | cleared | |||||||
received by the graph, never servedSHOPPING_ | excluded | excluded | excluded | excluded | excluded | excluded | excluded | excluded |
| Other products named | ||||||||
facts cleared for ad integrity checksADS_ | cleared | |||||||
Covid-sensitive Maps dataCOVID_ | cleared | |||||||
kid-friendly searchISOLATION_ | cleared | |||||||
In the schemas
Data bought under licence
Google buys from specialised suppliers the right to use their data: Gracenote sells TV and music metadata, Opta and Stats LLC sports statistics, Le Robert and Oxford their dictionaries. The supplier delivers its data continuously, as files or through an API: that is the feed. The contract sets what Google may do with it, and a flag carries that contract onto every fact.
CONTRACTUAL_NETFLIX_DATANetflix’s catalogueCONTRACTUAL_NETFLIX_GRACENOTEGracenote’s TV and music metadata on NetflixCONTRACTUAL_SPOTIFY_PODCAST_DATASpotify podcastsCONTRACTUAL_AUDIBLE_DATAAudible audiobooksCONTRACTUAL_STATS_LLC_DATAsports statistics from Stats LLC, next to Opta’sCONTRACTUAL_PRERELEASE_MUSIC_DATAinformation on an album before its releaseCONTRACTUAL_RANKING_PROVIDER_POPULARITY_DATApopularity scores bought from a ranking providerDICTIONARY_DATA_LE_ROBERTthe Le Robert dictionary, next to Oxford
What users and Google’s products contribute
The graph also receives what users say about themselves and what Google’s other products know: YouTube, Shopping, Maps and clicks in search.
JANATA_USER_GENERATED_DATApeople cards, profiles users create about themselvesSPII_KGOsensitive personal facts (SPII) supplied by the person who claimed their knowledge panelYOUTUBE_CHANNEL_ENTITIESYouTube channels turned into entitiesSHOPPING_PRODUCT_ATTRIBUTEa product’s attributes, with its description and imageSHOPPING_PRE_RELEASE_PRODUCTa product not yet releasedTECHNICAL_IN_DEVELOPMENT_GEO_KG_UGCMaps user contributions into the graph, in developmentINTENTJOINS_NB_SIGNALSprobably signals from NavBoost joined to search intents
Data written or computed by programs
Each type of machine-made data gets its own flag. Google thus decides, product by product, where that data may be used: film and TV synopses written by an LLM appear only in the products allowed to use them.
MINED_PEOPLEpeople spotted automatically in textsISOLATION_LLM_GENERATED_TVM_SYNOPSISfilm and TV synopses written by an LLMISOLATION_GENERATED_DATAmore broadly, automatically generated dataKE_TRUSTa trust score computed by the Knowledge EngineENTITY_TEXT_FEATUREStext features computed for each entityCIVICS_SEARCH_POLITICAL_TOPIC_SCORING_DATAa score that spots political topics, for searches on elections and institutions
Sensitive domains, kept in isolation
Flags that start with ISOLATION_ reserve a whole domain for the products cleared for it. Health, finance, horoscopes and kids’ search each have their own.
ISOLATION_MEDICALhealth, like finance (ISOLATION_FINANCIAL_DATA)ISOLATION_ASTROLOGY_HOROSCOPEhoroscopesISOLATION_KID_FRIENDLY_SEARCHkid-friendly searchISOLATION_S12Y_RECIPE_INGREDIENT_INSIGHTSthe environmental impact of recipe ingredients (S12Y stands for sustainability)ISOLATION_S3_REAL_ESTATEreal estate, like hotel roomsMOMA_RESTRICTED_TVCMOMA, Google’s intranet and its enterprise graph (employees, projects, documents), closed to temps, vendors and contractors (TVC)
Facts under a legal rule
LEGAL_UNSERVABLE_UNTIL_FURTHER_REVIEWa fact received but held until legal review: the only flag that stops a fact at the doorSPII_SEARCH_PUBLIC_INTERESTa sensitive personal fact served because the public interest justifies itACCC_RISKY_DATAdata at risk with the ACCC, Australia’s competition regulatorRIGHTS_MANAGEMENT_LLM_DATA_ACCESS_RESTRICTIONa restriction on LLM access to some data, on top of the AI restrictions every fact carriesLEGAL_PUBLIC_DOMAIN_USthe US public domain
Facts kept for the graph’s internal use
These facts stay inside the graph and feed its own computations.
TRIANGULATION_LOSERS_FOR_REFXfacts that failed triangulation, kept for RefXSPII_LOSERS_FOR_REFXrejected sensitive facts, kept the same wayISOLATION_PKG_*the personal graph (PKG, personal knowledge graph) of each Google account, in 85 compartments, one per source: Gmail bills, calendar, contacts, bookings… (83 start with ISOLATION_PKG_, 2 with PKG_)MATERIALIZED_ENRICHER_INTERNALfacts added by internal programs that complete the graph
The 187 requirements, sorted by reach
Follow the fact to the product that shows it 91
ISOLATION_22domains kept apart, reserved for cleared products
ASTROLOGY_horoscopesHOROSCOPE GENERATED_more broadly, automatically generated dataDATA KID_kid-friendly searchFRIENDLY_ SEARCH LLM_film and TV synopses written by an LLMGENERATED_ TVM_ SYNOPSIS MEDICALhealth, like finance (ISOLATION_FINANCIAL_DATA)S12Y_the environmental impact of recipe ingredients (S12Y stands for sustainability)RECIPE_ INGREDIENT_ INSIGHTS S3_real estate, like hotel roomsREAL_ ESTATE EXPERIMENT_ONLY FINANCIAL_DATA ONRAMP_DICTIONARY_ EXPERIMENT RWJ_CRUISE S3_CREATOR_ PRESENCE S3_CULTURAL S3_DATASEARCH S3_MEDIA_ ACTION S3_NLWEB S3_RECOMEDIA S3_TRAVEL_ HOTEL_ ROOMS SCOUTS_SOCIETAL_ CONTEXT_ ENTITIES WEB_CHANNELS YOUTUBE_MUSIC_ BASS_ ENTITIES YOUTUBE_OTT
CONTRACTUAL_24data bought under contract
ANDROID_Netflix’s catalogue, cleared for Android TVTV_ NETFLIX_ DATA AOG_food ordering through the AssistantFOOD_ ORDERING_ DATA ASSISTANT_the Assistant’s media actionsMEDIA_ ACTIONS AUDIBLE_Audible audiobooksDATA GOOGLE_Google Play filmsPLAY_ MOVIES_ DATA GTV_Google TV’s free channelsFREEPLAY_ LIVETV_ DATA NETFLIX_Netflix’s catalogueDATA NETFLIX_Gracenote’s TV and music metadata on NetflixGRACENOTE PRERELEASE_information on an album before its releaseMUSIC_ DATA RANKING_popularity scores bought from a ranking providerPROVIDER_ POPULARITY_ DATA SEARCH_Netflix’s catalogue, cleared for SearchNETFLIX_ DATA SPOTIFY_Spotify podcastsPODCAST_ DATA STATS_sports statistics from Stats LLC, next to Opta’sLLC_ DATA CARBON_INDEX_ APP CARBON_INDEX_ DATA IPG_DATA OPTA_DATA_ MIGRATION PRERELEASE_MUSIC_ DATA_ GPM PRERELEASE_YT_ SHOWS SLING_DATA STATS_DATA_ MIGRATION YETI_DATA YOUTUBE_ALC_ DATA YOUVIEW_DATA
SHOPPING_8Shopping data and where it may show
DO_received by the graph, never served to productsNOT_ PUBLISH_ TO_ TOPIC_ SERVER ON_only on google.comDOT_ COM_ ONLY ORGANIC_only in free listings, never in adsONLY PRE_a product not yet releasedRELEASE_ PRODUCT PRODUCT_a product’s attributes, with its description and imageATTRIBUTE ORGANIC_ON_ DOT_ COM_ ONLY PRODUCT_DESCRIPTION PRODUCT_IMAGE
RIGHTS_MANAGEMENT_6rights management
LLM_a restriction on LLM access to some data, on top of the AI restrictions every fact carriesDATA_ ACCESS_ RESTRICTION 100101102EBF_DATA INSIDER_RISK
MOMA_3Google’s intranet
RESTRICTED_MOMA, Google’s intranet and its enterprise graph (employees, projects, documents), closed to temps, vendors and contractors (TVC)TVC FTEUNRESTRICTED_TVC
Others28
ADS_facts cleared for ad integrity checksINTEGRITY_ ANNOTATION CIVICS_a score that spots political topics, for searches on elections and institutionsSEARCH_ POLITICAL_ TOPIC_ SCORING_ DATA DICTIONARY_the Le Robert dictionary, next to OxfordDATA_ LE_ ROBERT ENTITY_text features computed for each entityTEXT_ FEATURES INTENTJOINS_probably signals from NavBoost joined to search intentsNB_ SIGNALS JANATA_people cards, profiles users create about themselvesUSER_ GENERATED_ DATA KE_a trust score computed by the Knowledge EngineTRUST LEGAL_the US public domainPUBLIC_ DOMAIN_ US MINED_people spotted automatically in textsPEOPLE NETFLIX_the “watch on Netflix” buttonAVAILABILITY_ MEDIA_ ACTION TECHNICAL_Maps user contributions into the graph, in developmentIN_ DEVELOPMENT_ GEO_ KG_ UGC YOUTUBE_YouTube channels turned into entitiesCHANNEL_ ENTITIES YOUTUBE_YouTube podcastsPODCASTS_ DATA AOG_TEST_ DATA AR_ONBOARDING_ TEST AR_ONBOARDING_ TEST_ 2 COVID_MAPS_ SENSITIVE DICTIONARY_DATA_ OXFORD GEO_FEATURE_ RESTRICTION LIMITED_FOR_ SEMANTIC_ UNDERSTANDING LMS_POLICY_ ACKED_ ONLY MERCHANT_EVENTS_ DATA SENSITIVE_ENTITIES_ CLASSIFICATION SPII_WALDREF TECHNICAL_UNRECONCILED_ MEDIA_ ACTION TRAVEL_ACTIVITY_ DATA UMP_TESTING_ ONLY VIRTUALCARE_US
Keep the fact inside the graph, for its own computations 90
ISOLATION_PKG_83each Google account’s personal graph, one compartment per source
APP_NAMES ASSISTANT_CONTACT_ AFFINITY_ WITH_ METADATA ASSISTANT_DEVICE_ SETTINGS ASSISTANT_EPHEMERAL_ SUGGESTIONS ASSISTANT_LIST_ NAMES_ FOR_ SPEECH_ BIASING ASSISTANT_ROUTINES ASSISTANT_SETTINGS_ DEVICE_ SETTINGS ASSISTANT_SETTINGS_ FOOTPRINTS ASSISTANT_SETTINGS_ NICKNAME ASSISTANT_UPP_ PERSONAL_ TOP_ ENTITIES CALENDARS_FOR_ QUERY_ ANNOTATION CALENDARS_FOR_ SPEECH_ BIASING CALENDAR_EVENTS_ FOR_ QUERY_ ANNOTATION CALENDAR_EVENTS_ FOR_ SPEECH_ BIASING CONTACT_AGGREGATED_ SIGNALS CONTENT_INTERESTS DEVICE_INSTALLED_ APPS DISCOVER_SMART_ HOME_ DEVICES DYNAMIC_ENTITIES FLIGHT_LEG_ RESERVATIONS FOCUS_OWNER_ PROFILE FORMATTED_ADDRESS GAIAGELLER_ANSWERS GI_CAR_ RENTAL_ RESERVATION_ DATA GI_FLIGHT_ RESERVATION_ DATA GI_HOTEL_ RESERVATION_ DATA GI_RESTAURANT_ RESERVATION_ DATA GI_TRANSPORTATION_ RESERVATION_ DATA GMAIL_BILLS GMAIL_ORDERS HABITSHANDBAG_ENTITIES HERON_INTENTS_ AND_ TYPES HOTEL_RESERVATIONS HOUSEHOLDLAMS_PREFERENCES LOCAL_LEAF_ PAGE_ VIEW_ LOCATIONS LOCATION_SHARING_ CONTACTS MAPS_ALIAS_ FOOTPRINTS MAPS_SEARCH_ LOCATIONS MEDIA_HABITUAL_ CACHE MEDIA_LIBRARY MEDIA_USER_ CONTEXT_ INFO MEDIA_USER_ ENTITIES PARKING_LOCATIONS PEOPLE_API PEOPLE_API_ CONTACT_ ANNOTATIONS PEOPLE_API_ GET_ PEOPLE_ BY_ IDS PERSONALIZED_PRONUNCIATIONS PERSONAL_CONTACT_ ANNOTATIONS PERSONAL_PLACES PERSONAL_SHARED_ CONTACT_ ANNOTATIONS PLAY_AUDIO_ BOOKS PWS_ASSISTANT_ CONTACTS_ FOOTPRINTS RESTAURANT_RESERVATIONS SELF_ENTITIES SOCIAL_EVENT_ RESERVATIONS SOCIAL_GRAPH_ PEOPLE_ API STADIA_CONTACTS STARLIGHT_BULK_ LOOKUP STARLIGHT_BULK_ LOOKUP_ CONSISTENT STARLIGHT_BULK_ LOOKUP_ CONSISTENT_ CONTACTS STARLIGHT_COMPOSITE STARLIGHT_COMPOSITE_ FACE_ LABELS STARLIGHT_FACE_ LABELS STARLIGHT_QUERY STARLIGHT_TOP_ CONTACTS STARLIGHT_VISIBLE_ TO_ GUESTS STRUCTURED_MEMORY_ FOOTPRINTS TEACH_AND_ LEARN_ ENTITIES TYPED_QUERIES VANITY_COLLECTIONS VOICE_PROFILE WEBSEARCHWHITEPAGES_PHONE_ NUMBER YOUTUBE_ASSISTANT_ CONTEXT YOUTUBE_ASSISTANT_ XWALKSAFE_ CONTEXT YOUTUBE_MUSIC_ LOCKER YOUTUBE_PLAYLISTS YOUTUBE_PLAYLIST_ SEARCH YOUTUBE_PRIVATE_ PLAYLISTS YOUTUBE_PUBLIC_ AND_ PRIVATE_ PLAYLISTS
Others7
MATERIALIZED_facts added by internal programs that complete the graphENRICHER_ INTERNAL SPII_rejected sensitive facts, kept the same wayLOSERS_ FOR_ REFX TRIANGULATION_facts that failed triangulation, kept for RefXLOSERS_ FOR_ REFX ISOLATION_KE_ INTERNAL MATERIALIZED_SCUBED_ INTERNAL PKG_ASSISTANT_ CONTACT_ AFFINITY_ FOOTPRINTS PKG_RESERVATION_ DATA
Stay attached to sensitive facts Google displays 5
ACCC_data at risk with the ACCC, Australia’s competition regulatorRISKY_ DATA SPII_sensitive personal facts (SPII) supplied by the person who claimed their knowledge panelKGO SPII_a sensitive personal fact served because the public interest justifies itSEARCH_ PUBLIC_ INTEREST AU_ACCC_ RISKY_ FOR_ DISPLAY LEGACY_RESTRICTION
Stop the fact at the door 1
LEGAL_a fact received but held until legal review: the only flag that stops a fact at the doorUNSERVABLE_ UNTIL_ FURTHER_ REVIEW
The access requirement is set when the fact comes in. Livegraph’s record service turns it into a number, so a new requirement needs no update to the graph’s other systems: the list grows with every new supplier.
Every fact also carries a list of restrictions specific to generative AI: whether it may ground an answer in Search or outside Search, train a model, and whether a publisher’s Google-Extended opt-out applies. We detail them in the infographic on AI Overviews and AI Mode.
A restaurant changes its hours: four sources report it, Google keeps track of each
Take a restaurant that changes its opening hours, then closes for good. The information first goes through Google’s geographic database, the one behind Maps, before it reaches the Knowledge Graph, and every step records who said what.
- Step 1
Four sources report the change
The owner updates the Google Business Profile listing. Google can also call the restaurant with Duplex, its voice robot (Blue Ginger internally), one of whose uses,
OPENING_HOURS, checks opening hours. An extractor picks up the new hours from the restaurant’s website, and Maps contributors report the change. - Step 2
The Maps database keeps every source
Every place carries the list of its sources: the provider, one of the 800 the database knows, the collection date and, when the data comes from a page, the URL it was taken from. The place’s status takes four values: opening soon, open, temporarily closed, permanently closed.
- Step 3
The place becomes a graph fact
A Google component called protomesh converts the fields of the Maps database into graph facts, an operation the engineers call “triplification”. The closure becomes the fact
/geo/type/feature/permanently_closed, attached to the place’s key, its identifier in the graph:/geo/type/feature/id/{id}. - Step 4
Livegraph decides, TopicServer serves
The fact goes through Livegraph; when two sources disagree, its Composer decides. The chosen fact goes to TopicServer, which delivers it to Google’s products: knowledge panel, local pack, Gemini.
In the Maps database, every piece of data keeps the name of its provider, its collection date and the page it came from.
In the schemas
PROVIDER_GOOGLE_BLUE_GINGERGoogle’s automated calls to businessesPROVIDER_GOOGLE_LOCAL_WEB_EXTRACTIONextraction from business websitesPROVIDER_GOOGLE_GEO_UGC_TRUSTED_USERSMaps contributors deemed reliablePROVIDER_GOOGLE_GEO_MODERATIONmoderation of contributionsPROVIDER_GOOGLE_HUMEHume: a fix made in the graph flows back to Mapscollection_datethe date the data was collectedattribution_urlthe page the data comes from, “the ultimate source of the data (in case of scraping)”PERMANENTLY_CLOSEDpermanently closed, next to FUTURE_OPENING, OPEN and TEMPORARILY_CLOSED
Duplex’s phone calls also explore phone menus and check ad leads
Google counts the calls of Duplex, internally named Blue Ginger, among its ways of extracting facts for the graph. Among them: GEO_DATA_EXTRACTION (extracting place data), IVR_CRAWLING (exploring a business’s phone menu, the way a crawler explores a site), CHECK_WAIT (waiting time) and AD_LEAD_VERIFICATION (checking a lead from an ad). Publicly, Google says Maps receives 20 million contributions a day, filtered by machine learning and then audited by human teams.
At display time, the restaurant’s panel adds reviews, ratings from other sites and a summary written by Gemini
A knowledge panel shows more than stored facts: at display time, KEFI computes what is added to the graph’s facts.
The Wren & Thistle 1
Gastropub in York
4.6 (1,284)
Permanently closed 2
People mention: pie (37), beer garden (21), service (18) 3
Status confirmed by Google’s data 4
Reviews from the web food guide 9.1/10 · review site 4.5/5 6
Beer garden yes, according to the pub’s website 7
Address York 8
Fact stored in the graphComputed by KEFI at display time
- 1Name and type: graph facts from the Maps database, attached to the place’s key.
- 2Status: a graph fact written by Livegraph.
/geo/type/feature/permanently_closed - 3Summary of user reviews
UgcReviewSummaryResponse - 4Confirmation by Google’s own data
FirstPartyCorroborationResponse - 5AI-written summary: hung on the place’s MID.
LocalGenAISummariesResponse - 6The place’s rating on other sites
ThirdPartyRatingProto - 7Source backing a displayed attribute
SimpleAttributeCitationsResponse - 8Address: a graph fact from the Maps database.
The web, suppliers, editors and the entities themselves feed the graph
Google publishes the list in broad strokes: Freebase, Wikipedia and the CIA World Factbook at launch, in 2012, then “hundreds of sources from across the web”, licensed data, publishers’ structured data and feedback from the entities themselves. The schemas file every fact under one of four families, and give a separate place to people and brands that describe their own entity.
EXTRACTION
The web, read by extractors
Knowledge Vault reads text, HTML structure, tables and page annotations, and gives each fact a probability of being true. When Google Research presented it in 2014, it had extracted 1.6 billion facts, 271 million of them at 90% confidence or more; the live graph holds 500 billion facts. In the index, each page keeps the facts drawn from it, with a kgVerified flag when the graph already knows them.
PARTNER_FEED · THIRD_PARTY
Data bought from suppliers or sent by partners
THIRD_PARTY covers data bought under licence: for TV, music, sports, audiobooks and dictionaries, Google buys from specialised suppliers such as Gracenote or Opta. PARTNER_FEED covers feeds a partner sends itself in a format Google sets, such as the catalogue a video service describes in a Media Actions feed so its films show in Search and the Assistant. Every fact received carries its contract’s flag. Shopping, YouTube and the Maps database feed the graph too, each on its own terms: some Shopping products enter it without ever being served.
CURATION
Human editors
The court ruling speaks of “thousands of employees and contractors”. They work in Hume. A schema team owns the ontology, the list of entity types and their properties: a person has a birth date, a film has actors.
A place of their own
The entities themselves
A person or a brand can claim its knowledge panel, that is get Google to recognise it as the entity described. It then becomes an authority on itself: its facts have a status of their own, and each one traces back to the request it came from. People cards, profiles users create about themselves, launched in India in 2020 under the internal name Janata (“the people” in Hindi), go through a ten-state moderation flow, from STATE_DRAFT to STATE_REJECTED_FOR_IMPERSONATION.
Every entity has its reference pages
For each entity, Google keeps pages that are “highly topical for this entity” and uses them to mine further facts. The 2024 leak’s examples: for Apple, its Wikipedia page and apple.com; for an actor or a film, its IMDb page. One score picks these pages (referencePageScore), another measures whether a page is about this entity only (singleTopicness), and one field records for which queries, in which country and language a page is the official one (queriesForWhichOfficial).
Filling the gaps by asking the search engine
“Over 70% of people included in Freebase have no known place of birth”, Google researchers wrote in 2014. Their method learns, for each attribute, which queries to ask the search engine so that the snippets contain the right value. Biperpedia, an ontology published by Google Research the same year, mined 1.6 million (class, attribute) pairs from queries and the web: what people search about a type of entity becomes a property to fill.
An entity’s names come from 21 sources: Wikipedia, its anchors and clicks
An entity keeps each of its names with its origin, among 21 sources. Wikipedia supplies four: the article title, its redirects, the text of the links pointing to it (anchors) and disambiguation pages. Freebase, the WordNet lexical database, Oxford and Google Play add theirs. NavBoost adds the words users type before clicking on the entity’s pages. The 2024 leak gives the trust rule: a name counts above 0.6, with a score drawn “from quantitative sources such as navboost clicks, anchors”.
Elizabeth Reid, head of Google Search, testified that click and query data does not play a “particularly meaningful” role in building the graph. The schemas put them on the recognition side: they teach Google the names people search an entity under, and facts come from the sources.
Some of the 21 name sources
SRC_WIKIPEDIA_NAMEthe Wikipedia article titleSRC_WIKIPEDIA_REDIRECTits redirectsSRC_WIKIPEDIA_ANCHORthe text of links to the articleSRC_WIKIPEDIA_DISAMBIGdisambiguation pagesSRC_NAVBOOSTqueries followed by clicks on the entity’s pagesSRC_OXFORDthe Oxford dictionaryvolumeBasedScorethe score drawn from clicks and anchors
Livegraph attaches each fact to the right entity before writing it
Livegraph is the Knowledge Graph’s write pipeline: every incoming fact goes through it, in six steps, from the feed that sends it to the server that delivers it to Google’s products.
- InputThe feedA provider sends its facts, each with its passport.
- CheckRecord serviceThe record service checks the format and turns the access requirement into a number.
- ReconciliationWhich entity?Attaches the fact to the right key, or to a new entity.
- ArbitrationComposerPicks the value to keep when sources disagree.
- InferenceFactStoreHolds the composed facts and those inferred from them.
- ServingTopicServerDelivers entities to products, from the knowledge panel to Gemini.
Reconciliation happens up front, as the fact comes in
Reconciliation means deciding which entity a fact is about: is this “Apple” the brand or the fruit? The Google engineers who designed schema.org put it plainly: the Knowledge Graph is distinguished by “a strong emphasis on up-front entity reconciliation”. A feed can send a phone number or an address for the sole purpose of helping Google find the right entity: those clues carry the is_supporting_data flag and stay out of the graph that products see.
Reconciliation relies on numeric thresholds. To match a TV operator’s catalogue with graph entities, four tolerances are used:
In the schemas
title_edit_distancethe tolerated gap between two titles, in characterstitle_edit_distance_percentagethe same gap, as a share of the titlerelease_year_diffthe tolerated gap between two release yearsduration_minute_diffthe tolerated gap between two running times, in minutes
How strict reconciliation is depends on the domain
Brands are “fully reconciled with KG entities so there should never be ambiguity”, while Shopping’s data models “do not always line up 1-1 with KG entities”. Duplicate places point to the one that replaces them (duplicate_of), and a local knowledge panel is tagged merged or not (LOCAL_KP_MERGE, LOCAL_KP_NONMERGE).
For a product, Google separates three levels: the model (a /g/ key), the item sold, identified by its GTIN, the barcode (a /t/ key), and a merchant’s offer.
Google Cloud sells a slice of this engine: its Entity Reconciliation API “wraps around the Google core entity resolution engine”, matching on fuzzy text, common relationships, entity types and attributes.
8,679 collections sort the graph’s entities and decide where Google can offer them in a request
A graph entity gets into a request only through its collections, sets of entities of the same kind: /collection/films, /collection/brands, /collection/bakery_businesses. Take “play Oppenheimer”, “book a table for Christmas” and “pay with my card”: each request has a slot to fill (a film, an occasion, a payment method), and only entities from the collections that open that slot can fill it. A table from the Assistant, Google’s voice assistant whose actions Gemini takes over on phones, links 11,637 graph identifiers to 397 slots: 8,679 collections, and 2,958 entities named one by one when the list is closed. For a brand, the right collection decides whether Google can offer it.
Products, people, occasions: what each family of collections lets users ask for
One collection often opens several slots: the 140 drug classes are both products to buy and allergens, a film is also the topic of a quiz.
3,261 collections · slot Commercial_product_type
Products to buy
the product type of a shopping request, including 500 product lines
/collection/ bar_ stools /collection/ rugs /collection/ mulch_ products /collection/ woks_ product_ lines /collection/ air_ purifiers_ product_ lines
870 collections · slot Person
People
a person, by trade, by the language they speak or by what they studied
/collection/ blacksmiths /collection/ curators /collection/ punjabi_ speakers /collection/ medicine_ majors
777 collections · slot Organization
Organisations
323 industries and 210 kinds of associations and bodies
/collection/ bakery_ businesses /collection/ venture_ capital_ businesses /collection/ malaria_ organizations /collection/ rugby_ union_ leagues
674 collections · slot Music_person
Music artists
an artist to listen to, filed by genre
/collection/ tango_ artists /collection/ americana_ artists /collection/ show_ tune_ artists /collection/ alternative_ rock_ artists
604 collections · slot Reservation_occasion
Occasions
the occasion of a booking: holidays and hundreds of sports competitions
/collection/ holidays /collection/ ryder_ cups /collection/ 24_ hours_ of_ le_ mans /collection/ 2000_ guineas_ stakes
545 collections · slot Album_title
Albums
an album, by music genre
/collection/ teen_ pop_ albums /collection/ new_ romantic_ albums /collection/ grunge_ albums /collection/ soca_ albums
325 collections · slot Film_title
Films
a film to watch, but also the theme of a greeting card or the topic of a quiz
/collection/ black_ comedy_ films /collection/ caper_ story_ films /collection/ 16_ mm_ films /collection/ 3-d_ films
295 collections · slot Account_provider
Account-based services
software, a game, a site or a supplier where the user has an account
/collection/ mmorpg_ video_ games /collection/ integrated_ development_ environment_ software /collection/ fansite_ websites /collection/ electric_ services_ businesses
210 collections · slot Athlete
Athletes
an athlete, by sport
/collection/ olympic_ artistic_ gymnasts /collection/ professional_ trampoline_ athletes /collection/ brazilian_ jiu-jitsu_ martial_ artists
194 collections · slot Book_title
Books
a book, by genre or format
/collection/ historical_ novel_ books /collection/ almanacs /collection/ guidebooks /collection/ e-books
150 collections · slot Allergen
Allergens
what someone is allergic to: 140 drug classes, chemical compounds, domestic animals
/collection/ opioid_ drugs /collection/ over-the-counter_ drugs /collection/ amino_ acids /collection/ domesticated_ animals
141 collections · slot TV_program
TV programmes
a show or a series, by genre
/collection/ tv_ medical_ dramas /collection/ christmas_ tv_ specials /collection/ tv_ variety_ shows
The brand, the company and the product line each get in through their own collections
The brand slot (Brand) opens through only six collections: /collection/brands, /collection/shopping_brands, /collection/fashion_labels, /collection/drug_brands, /collection/foods_and_food_products and /collection/business_operations. A fashion house filed only under an industry stays an organisation. Take a cosmetics group: three entities, three collections.
- The company
/→collection/ cosmetics_ businesses Organizationrecognised as an organisation - Its brand
/→collection/ brands Brandrecognised as a brand - Its product lines
/→collection/ cosmetics_ product_ lines Commercial_product_typerecognised as a product to buy
A company’s industry decides which actions it can serve: paying, travelling, keeping informed
Out of 347 industry collections, 323 lead to the Organization slot alone. The other 24 lead to an action. A commercial bank becomes a source of money to pay with (Money_source), a pharmacy fills the Pharmacy slot, a rail company the trip slot (Transit_provider), a news agency the news provider slot, an electricity company the account where a bill gets paid.
The exact collection matters: /collection/commercial_bank_businesses opens Money_source, its near-namesake /collection/commercial_banking_businesses only leads to Organization.
Collection names reuse Google Shopping’s categories and industry classifications
Nine collections, under the merlot/ prefix, carry the number of a category from the Google product taxonomy, the one merchants fill in Merchant Center (google_product_category): merlot/makeup_477 is category 477, Makeup. For these nine categories, the category a merchant gives a product matches a graph collection.
The 9 merlot/ collections and their Google category
merlot/electronics_222Electronics (222)merlot/media_streaming_devices_5276… > Streaming & Home Media Players (5276)merlot/health_and_beauty_469Health & Beauty (469)merlot/cosmetics_473… > Personal Care > Cosmetics (473)merlot/makeup_477… > Cosmetics > Makeup (477)merlot/food_beverages_and_tobacco_412Food, Beverages & Tobacco (412)merlot/beverages_413… > Beverages (413)merlot/tea_and_infusions_2073… > Beverages > Tea & Infusions (2073)merlot/toys_and_games_1239Toys & Games (1239)
Industries often reuse the titles of the US SIC and NAICS classifications: perfumes_cosmetics_and_other_toilet_preparations_businesses is SIC code 2844. 17 collections come from the Maps places database: store chains, restaurants, transit lines.
Only entities with “authority potential” can be cited as the source of an FAQ
Take a request such as “what does my bank’s FAQ say about transfers?”: the bank fills the FAQ source slot (FAQ_provider), which only opens to entities in /collection/knowledge_panels/authority_potential. For a brand, this collection decides whether Google can name it as the source of an answer. A whitelist of official websites likewise opens the slot of services where the user has an account. Both belong to the 17 collections Google maintains for its knowledge panels.
…/authority_potentialthe only collection that opens the FAQ source slot (FAQ_provider)…/official_websites_whitelista whitelist of official websites, which opens the slot of a service where one has an account…/fresh_content_sourcesfresh content sources, which open the news provider slot…/shopping_kp_eligibleproducts eligible for the Shopping knowledge panel…/social_media_providerssocial networks…/youtube_creatorsYouTube creators, beside channels and their owners
Each collection carries two scores: what the entity is known for, and its weight against the other members
Google scores each of an entity’s collections twice. notable_for_score: is this actor turned governor an actor first? notable_within_score: among actors, how much weight does this one carry? A star weighs more than a supporting player. In Google Search, the knowledge panel subtitle (“Actor”, “Clothing brand”) shows what the entity is known for. SAFT turns them into two probabilities: the collection given the entity, and the entity given the collection.
In the schemas
notable_for_scorehow much the entity is known as a member of this collectionnotable_within_scorethe entity’s weight among the collection’s membersnum_notable_typesthe number of types the entity is known asCOLLECTION_GIVEN_ENTITYthe collection given the entityENTITY_GIVEN_COLLECTIONthe entity given the collection
When the list is short, Google names each entity: 54 pizza toppings, 62 payment methods
In all, 2,958 entities are named one by one, for 248 slots. Other slots open through whole collections, such as destroyers or cranial nerves.
- 54
pizza toppings
Pineapple, ranch dressing, paneer, Buffalo wings, and Pizza Margherita itself, as a topping.
The 54 pizza toppings
- Anchovy
- Artichoke
- Asparagus
- Bacon
- Banana pepper
- Barbecue
- Barbecue chicken
- Basil
- Beef
- Bell pepper
- Bolognese sauce
- Buffalo wings
- Capers
- Cheddar cheese
- Cheese
- Cheesesteak
- Chicken and mushroom pie
- Clam
- Feta
- Fish
- Garlic
- Goat cheese
- Gorgonzola
- Ham
- Hawaiian pizza
- Hot sauce
- Italian sausage
- Jalapeño
- Leek
- Meat
- Meatball
- Mozzarella
- Mushroom
- Olive
- Onion
- Paneer
- Parmesan
- Peperonata
- Pepperoni
- Pesto
- Pineapple
- Pizza Margherita
- Prawn
- Ranch dressing
- Red onion
- Rocket
- Rosemary
- Salami
- Sausage
- Seafood
- Spinach
- Sun-dried tomato
- Tomato
- Wisconsin cheese
- 93
coffee-shop drinks
From espresso to the Unicorn Frappuccino, sold for five days in April 2017. Separately, 58 customisations, from oat milk to Sweet’n Low.
- 62
payment methods
Cheques, Apple Pay, Venmo, Messenger and Dogecoin, side by side.
- 107
securities
Including Samurai, Panda, Kimchi and Baklava bonds.
- 17
dice games
Dungeons & Dragons next to Yahtzee and backgammon.
The body part to work out
Muscles, but also veins, arteries and cranial nerves.
The thing whose sound someone asks for
An instrument, an animal, a submarine, an aircraft carrier or a destroyer.
Quiz
True or false, for Google?
Pizza Margherita is a pizza topping.
True. It is one of the 54 toppings, like Hawaiian pizza.
Cheques have dropped out of the payment methods.
False. They are among the 62, next to Apple Pay, Venmo and Dogecoin.
A drug class can be an allergen.
True. The 140 drug classes are both products to buy and allergens.
A business in
/collection/commercial_banking_businessescan be used to pay.False. Only its near-namesake
commercial_bank_businessesopens theMoney_sourceslot;commercial_banking_businessesonly leads toOrganization.The 24 Hours of Le Mans is a booking occasion.
True.
/collection/24_hours_of_le_mansopens theReservation_occasionslot, like holidays and hundreds of competitions.
When two sources disagree, the Composer decides using three settings
When two sources give two different values for the same property (a birth date, a name, opening hours), Livegraph’s Composer decides. Three settings steer its choice.
Weak data comes after the rest
Data flagged weak_data loses against any regular data. A new feed thus enters the graph without breaking anything: it only shows when nothing contradicts it.
A priority per slice of a feed
A ranking_token gives part of a feed its own priority. The engineers’ example: the Maps “geo” feed can put some of its place names ahead of other sources’.
Three different sites
A fact flagged requires_triangulation must arrive from three distinct sites. A fact that arrives once citing three sites counts as one confirmation: it “cannot self-triangulate”.
Ten pages on one site count as a single source
Take a fictional brand, Hollis & Grey, and a fact that came through a feed subject to triangulation: “Hollis & Grey, founded in 1998”. Tick the pages that repeat it.
Triangulation is a setting each feed switches on
The engineers address “a new client trying to enable triangulation for your feed”: each team that sends facts to the graph chooses whether to switch it on. Because the rule counts sites, it targets facts extracted from the web.
Google’s patents also count domains, not pages. The one on discovering new entities discards clusters seen on too few domains (US9785696B1); another lowers the corroboration threshold for documents whose facts were already corroborated (US8682913B1). In 2015, Knowledge-Based Trust, a method published by Google Research, already judged a source by the accuracy of its facts: 2.8 billion extracted facts, 119 million pages evaluated.
Rejected facts stay in the graph
A fact that fails triangulation stays stored with the TRIANGULATION_LOSERS_FOR_REFX flag (losers, in the engineers’ words). Google does not display it, but RefX uses it. Rejected sensitive facts follow the same path (SPII_LOSERS_FOR_REFX).
The FactStore, inside Livegraph, also infers facts: a person born in Lyon was born in France. When a source asserts a fact the graph has already inferred, its provenance is added to the existing fact without rewriting it (provenance_only_addition): the inference gains one more source.
The graph keeps track of what it knows is false, what it does not know and what it deletes
A correction adds information to the Knowledge Graph: the fact is marked false, unknown, deleted or merged with a duplicate, and the correction keeps its author and date.
In the schemas
is_negationthe fact states that something is false: the graph keeps “what we know to be true and what we know to be false”HAS_UNKNOWN_VALUEthe value exists but Google does not know it, like an unknown date of deathHAS_NO_VALUEthere is no value: no date of death, the person is alivedeleted_onthe date a fact was deleteddeletion_provenancethe deletion’s passport: who erased the factduplicate_ofa duplicate place points to the one that replaces itlegal_removal_regionsthe countries where a fact must be removed after a legal requestavailability_end_timestampthe date after which a fact can no longer be used; its counterpart, the start date, sets an embargo
A merged entity keeps its old keys
When Google merges two duplicate entities, both keys keep working. The 2024 leak says so for WebRef: the old ids “represent the exact same entity as the one they were merged with”. One key seen in a query and the other in a page mean the same thing. Google patented this forwarding in 2006 (US7991797B2).
A sensitive personal fact can be certified in only three ways
The fact comes from the person themselves, from a legal request or from public information. For the last one, the engineers specify:
Engineers’ commentThis fact is public information. (See go/kg-spii-certification for a description of what qualifies as public information -- simply finding a fact online is not sufficient to certify a fact as public.)
authority_feedbackthe fact comes from the entity itself, owner of its panellegal_requestthe fact comes from a legal requestpublic_informationthe fact is public information, recognised as such
Publicly, Google may remove information that is “demonstrably false or outdated”, with evidence: legal documents, expert consensus or reputable primary sources. It does not create or delete knowledge panels by hand: panels appear automatically “when there is enough information available on the open web”. Upkeep also comes in large clean-ups: Kalicube, an agency specialised in knowledge panels, measured in June 2025 a 6.26% drop in entities, and about 77% for Event entities.
The graph hands out hundreds of millions of numbers a day, and their order dates each entity’s arrival
Every entity gets a key, which Google calls a MID (machine ID): /m/ for those inherited from Freebase (bought in 2010), /g/ for the others. /g/ keys come out of a counter, in writing order. By placing 26,815 dated entities along that counter, our kg-mid-dating tool dates a key’s entry into the graph within three months in 85% to 96% of cases depending on the key range, with a median error of five to eight days.
The counter numbers every object in the graph, well beyond visible entities: about 4 million a day in 2015-2016, 15 million from 2017 to 2020, then 30 to 270 million.
Since you opened this page, the counter has handed out0numbers, at the 2026 pace: about 210 million a day.
An unpredictable event enters the graph within days
The Notre-Dame de Paris fire (15 April 2019) got its key, /g/11h10vl01n, the same month, as did the assassination of Shinzō Abe (8 July 2022); the Myanmar earthquake (28 March 2025), the following month. Kalicube saw online events often become entities “within 15 minutes” of being published, then disappear shortly after.
Films, games and elections enter as soon as they are announced; organisations later, six in ten more than three months after they were founded. A key also outlives a change of name: Gemini’s, /g/11ts49p01g, dates from January 2023, when Google was preparing Bard, renamed Gemini a year later.
How an entity is born
On the web: one patent (US9785696B1) merges graphs extracted from similar pages, discards those seen on too few domains and outputs “a suggested new entity and entity relationships”; another (US10198491B1) treats a word sequence missing from known entities as a new entity candidate.
Through users: people cards, which they create about themselves, go through phone, identity and content checks.
Date an entity’s entry into the graph
Paste a MID (/g/11…) or the address of a Google results page that contains kgmid=: that is the case when you open an entity’s knowledge panel.
Examples:
Computed in your browser from 4,470 reference entities: within three months in 8 cases out of 10. The full tool, kg-mid-dating, gives the month and its range.
Knowledge Graph size announced by Google. The jump from 70 to 500 billion facts between 2016 and 2020 also reflects how facts are counted.
A club for entities. Join the list: you get a member number, your MID and a link of your own. Nothing else is promised.
For SEO: a fact wins when independent sites and an authoritative source carry it
The Knowledge Graph feeds knowledge panels, direct answers, AI Mode, AI Overviews and Gemini. The way it chooses and files its facts has ten practical consequences for a brand or a business.
Mentions on different sites
When a feed requires triangulation, a fact must arrive through three distinct sites: ten pages on one site count as one. The facts that matter for a brand, founder, founding date, headquarters, products, gain from being picked up by independent sites: press, industry directories, partners, Wikipedia and Wikidata.
An official page focused on the entity
Google picks reference pages for each entity, with a score that measures whether a page is about that entity only. A complete “About” page that describes the brand and nothing else is the natural candidate, like apple.com for Apple.
Structured data that helps reconciliation
Google’s schema.org authors describe these annotations as a Knowledge Graph source for logo, contact and social information, and ask publishers to describe each entity well enough for reconciliation. Organization markup, with sameAs pointing to official profiles, serves that purpose first.
Claim your knowledge panel
A verified entity is an authority on itself: its feedback is handled first, and the facts it supplies have a status of their own in the passport. Google traces each of these facts back to the original request.
For a business, the Maps listing is the way into the graph
Hours, status and address flow from the Maps database into the Knowledge Graph. Google cross-checks the listing, the website, contributions and, sometimes, a phone call: identical information everywhere keeps one source from contradicting another. Reviews and ratings on other sites also show in the panel, computed at display time.
Consistent anchors and brand searches
An entity takes its names from Wikipedia titles and anchors, and from queries followed by clicks on its pages. Links that name the brand the same way, and brand searches that lead to the site, teach Google the names to recognise it by.
The same GTINs at every retailer
Google identifies a product for sale at barcode level. Clean GTINs, repeated identically by every retailer, make a product the same entity everywhere, in Shopping as in AI Mode.
Know your key and its date
The
kgmidparameter of a results page gives an entity’s key; the dating tool above gives its entry into the graph. A brand that entered late, or exists under two keys, shows up at once. When Google merges two keys, both stay valid.Get filed as a brand
Only six collections open the brand slot in a request,
/collection/brandsfirst, and the knowledge panel subtitle shows the type Google keeps. The levers: a Wikidata item that states what the brand is (instance of: brand, fashion label) and its industry, Organization structured data linked to that item with sameAs, with the NAICS code (naics property), the brand property on product pages and the right Google category in Merchant Center.Correct with evidence
Google removes what is “demonstrably false or outdated”, with evidence. A request backed by a legal document or a primary source has its place in the fact’s passport; a plain assertion does not.