A knowledge graph is a machine readable map of entities (people, companies, places, products, concepts) and the verified relationships between them, and in 2026 it is the layer that decides whether ChatGPT, Perplexity, Gemini, Claude, and Google AI Overviews can name your brand at all. Google’s Knowledge Graph held more than 1.6 trillion facts about 54 billion entities as of May 2024, up from 500 billion facts about 5 billion entities in 2020, and Wikidata created its 140 millionth item in May 2026 while holding roughly 122 million live items. If your business is not a resolvable node inside those graphs, and inside the graph an AI engine builds at query time, the model has nothing stable to attach a recommendation to.
That is the part most coverage of this topic skips. Every guide explains that a knowledge graph stores subject, predicate, object triples. Almost none explain the three graphs that actually govern AI citations, why Google deleted more than 3 billion entities from its own graph in June 2025, or why Microsoft Research built GraphRAG to construct a graph on the fly because the vector index alone kept failing. Here is the working version, written for people who need their brand in the answer.
What is a knowledge graph and how is it different from a search index?
A search index stores documents and the words inside them. A knowledge graph stores things and the facts that connect them. The index knows that a page contains the string “Hendricks Law Firm.” The graph knows that Hendricks Law Firm is an Organization, founded in a specific year, located in a specific city, practicing a specific area of law, led by a named attorney, and identical to a specific LinkedIn page and Wikidata item.
That difference matters because large language models generate answers about entities, not about strings. When a user asks ChatGPT for the best estate planning attorney in Charleston, the model is not scanning your homepage word by word. It is assembling a shortlist of entities it can resolve with confidence, then checking what corroborated facts it holds about each one. Schema.org markup delivered as JSON-LD is how you hand a machine the entity version of your business rather than the string version. That is the whole discipline behind entity SEO for AI search, and the graph is where the work lands.
Which knowledge graphs decide whether AI names your brand?
Five separate graphs matter, and most brands only ever work on one. Treat them as distinct systems with distinct entry points:
1. The Google Knowledge Graph. The canonical entity store behind Knowledge Panels, AI Overviews, and Google AI Mode. Gemini is trained on and grounded against it, so your representation here influences answers on the largest surface in search.
2. The open graph: Wikidata, Wikipedia, and DBpedia. Wikidata is the open, editable, structured database that feeds nearly every downstream system. DBpedia extracts structured data out of Wikipedia articles. These are the graphs a startup can actually get into, because they accept submissions rather than waiting to discover you.
3. The model’s internal graph. The entity relationships baked into model weights during training. Slow to change, updated only when a model retrains, and shaped almost entirely by how often your brand appeared in the training corpus with consistent attributes.
4. The retrieval time graph. What systems like Microsoft GraphRAG build at query time by extracting entities and relationships from retrieved documents, clustering them into communities, and summarizing each cluster before answering.
5. Your own site graph. The Organization, Person, Service, and Article schema on your domain, linked together by @id references and pointed outward with sameAs. This is the only one of the five you fully control.
Want to know which of those five graphs currently contain your brand and which return nothing? Run a free AI visibility audit and get the entity map for your domain, your competitors, and the queries your buyers actually type.
Why does AI search need a knowledge graph if it already has vector search?
Because vector search retrieves things that sound similar, and graphs retrieve things that are actually connected. A vector index converts text into embeddings and returns the passages nearest your query in that space. It is fast and it scales, and it is genuinely good at fuzzy topical recall. It is bad at multi hop reasoning, at questions that span documents, and at telling two entities with similar names apart.
Microsoft Research shipped GraphRAG for exactly that reason. Instead of retrieving isolated chunks, GraphRAG uses an LLM to extract an entity and relationship graph from the source corpus, detects communities of densely connected nodes at multiple levels, generates a summary for each community, and then answers in one of two modes: global search across community summaries for corpus wide questions, and local search that fans out from a specific entity to its neighbors. That structure is what lets a system answer “which firms handle both immigration and employment cases in this metro” instead of just returning the page that used those words most.
The measured gains are real but not magic. Lettria’s benchmarking found graph based retrieval improved precision by as much as 35% over vector only pipelines, and comparative testing has shown hybrid GraphRAG improving factual correctness by roughly 8% and context relevance by roughly 7% against standard retrieval augmented generation. Most production systems now run both: vectors for broad recall, graph traversal for structured reasoning and explainability.
How did Google build its Knowledge Graph, and what changed in 2025?
Google bought the graph. It acquired Metaweb, the company behind Freebase, on July 16, 2010, and Freebase became a named primary source when the Knowledge Graph launched in 2012. Google announced Freebase’s shutdown in December 2014, closed it to edits, and finally shut it down on May 2, 2016, after mapping roughly 4.6 million topics into Wikidata and generating about 14 million new statements in the process. That migration is why Wikidata sits so close to the center of entity visibility today.
Then 2025 broke the assumption that the graph only grows. In June 2025 the Google Knowledge Graph recorded its largest contraction in a decade, a two stage drop of 6.26% inside a single week that removed more than 3 billion entities. Entities carrying the generic “thing” label fell by about 15.27%, roughly 8 billion of them. Read plainly, Google purged entities it could not confidently type or corroborate.
The practical lesson for any business is uncomfortable and useful. Being present in the graph is not permanent. An entity with a thin, unverifiable, weakly typed profile is exactly the kind of node that gets pruned, and a pruned entity stops feeding Knowledge Panels, AI Overviews, and Gemini grounding. Corroboration across independent sources is what keeps you resident. That is the same logic that governs whether you earn a Google Knowledge Panel for AI search.
How do you build an entity footprint AI engines can resolve?
Give every major system the same set of facts, then connect those systems to each other. Entity resolution is the process of deciding that two records describe the same real world thing, and you win it by removing ambiguity rather than by adding volume.
Start with Organization schema in JSON-LD on your homepage: legal name, logo, founding date, address, phone, area served, and the services you sell. Add Person schema for your founders and named experts with real credentials. Then add sameAs, the Schema.org property that asserts two representations point at the same entity. Priority order for sameAs targets: your Wikidata Q identifier first because it feeds Google directly, then Wikipedia if you qualify, then LinkedIn, Crunchbase, your Google Business Profile, industry directories, and your verified social profiles. A well established brand typically carries eight to twelve sameAs entries, and each one lowers the odds that a machine treats your two profiles as two companies.
Wikidata deserves its own push because it is the strongest entry point available to a business that is not yet famous. Items need a label, a description, an instance of statement, and referenced claims sourced to independent coverage. Wikipedia is a different threshold entirely: the General Notability Guideline requires significant coverage in reliable sources independent of the subject, and the organizations and companies standard is stricter still, explicitly discounting routine product announcements, funding notices, and press releases. Most companies do not qualify, and submitting anyway earns a deletion nomination. The workaround is not a trick, it is earning genuine third party coverage first. Our breakdown of Wikidata for AI visibility covers the submission mechanics step by step.
What breaks entity resolution, and how do you fix it?
Four failures cause most of the damage, and all four are repairable inside a quarter. First, name drift: the LLC name on your filings, the name on your Google Business Profile, the name in your footer, and the name your PR uses are all slightly different, so machines create separate nodes. Pick one canonical string and enforce it everywhere.
Second, address and phone inconsistency across directories, which splits a single location into competing records. Third, duplicate or abandoned profiles, including old office listings, a second LinkedIn page from a rebrand, and a Crunchbase entry with a defunct URL. Fourth, unlinked islands: strong profiles on LinkedIn and Crunchbase that never point at each other or back at your site, leaving nothing for a resolver to follow.
There is also a tooling change worth knowing. Microsoft retired the entire Bing Search API family, including the Entity Search API, on August 11, 2025, and pointed developers to Grounding with Bing Search inside Azure AI Agents. That replacement is an agent product rather than a structured results endpoint, so the era of cheaply querying a public entity API and reading your record back is over. Verification now happens through the engines themselves: query ChatGPT, Perplexity, Claude, Gemini, and Microsoft Copilot directly and read what they return about you.
How do you measure whether knowledge graph work is landing?
Track resolution, not rankings. The four metrics that matter are: does a Knowledge Panel appear for your brand name on Google, does your Wikidata item exist and carry referenced claims, do ChatGPT and Perplexity describe your business accurately when asked directly, and do they include you in the category and location prompts your buyers use.
Run those checks monthly and log the answers. Entity work is slow and then sudden. The Knowledge Panel usually shows up weeks after the underlying corroboration lands, not the day you publish schema, and model level recognition lags further because it depends on retraining cycles. Watch for regressions too, because the June 2025 purge proved that graph presence can be revoked. The brands that hold their position are the ones treating entity data as infrastructure they maintain, not a project they finished.
Frequently asked questions
What is a knowledge graph in simple terms? It is a database of things and the relationships between them, stored as facts a machine can traverse. Instead of documents containing words, it holds entities like companies, people, and places connected by typed relationships such as founder, located in, or same as. The Google Knowledge Graph and Wikidata are the two best known public examples, and both feed the answers ChatGPT, Gemini, and Perplexity produce.
How big is the Google Knowledge Graph in 2026? Google reported more than 1.6 trillion facts about 54 billion entities as of May 2024, up from 500 billion facts about 5 billion entities in 2020. That trajectory reversed in June 2025, when the graph shed 6.26% in one week, removing over 3 billion entities, with generically typed “thing” entities dropping about 15.27%. Size alone is not the story: corroboration quality decides who survives a purge.
Is a knowledge graph the same as vector search? No. Vector search finds passages that are semantically similar to a query using embeddings. A knowledge graph stores explicit, typed relationships you can traverse and audit. Graph retrieval improved precision by up to 35% over vector only pipelines in Lettria’s benchmarking, and Microsoft GraphRAG combines both. Production AI search systems generally run hybrid: vectors for recall, graphs for reasoning and explainability.
Do I need a Wikipedia page to appear in AI answers? No, and most businesses will not qualify. Wikipedia’s General Notability Guideline requires significant coverage in independent reliable sources, and the standard for companies explicitly excludes routine announcements and press releases. Wikidata has no notability bar that high and feeds Google directly, so it is the practical entry point. ChatGPT cites Wikipedia for around 12% of its citations, but Perplexity barely uses it at all.
What is sameAs and why does it matter for AI search? sameAs is a Schema.org property asserting that two web representations describe the same real world entity. Adding sameAs links from your Organization schema to your Wikidata item, LinkedIn page, Crunchbase profile, and Google Business Profile lets resolvers merge those records into one confident node instead of several ambiguous ones. Established brands typically carry eight to twelve sameAs entries, ordered with Wikidata first.
How long does it take to build an entity footprint? Expect 60 to 120 days before a Knowledge Panel or accurate AI description appears, assuming you fix name and address consistency, ship Organization and Person schema, create a referenced Wikidata item, and earn independent coverage in that window. Model level recognition inside ChatGPT and Claude takes longer because it depends on retraining. The work compounds, and the sequence matters more than the speed.
Search used to reward the site that best matched a phrase. AI search rewards the entity a system can identify without hesitating, because a model will not stake an answer on a company it cannot resolve. Your competitors are already showing up as clean nodes with a Q identifier, a Knowledge Panel, corroborated facts, and a dozen sameAs links pointing home. Every month you stay ambiguous is a month the engines quietly recommend someone else by default.
Ready to stop guessing whether the engines know who you are? Request your entity and citation audit and we will map your presence across the Google Knowledge Graph, Wikidata, and the major AI platforms, then hand you the fix list in priority order.
Tagged