Entity Grounding and the Knowledge Graph

Three different operations, one name

An entity is a thing with an identity — a person, a company, a place, a product, an event — as distinct from the string of characters used to name it. Google's framing when it launched the Knowledge Graph in 2012 was "things, not strings." Entity grounding, also called entity linking or entity resolution, is the step in which a system reads a name in text and decides which known thing it refers to: that this "Apple" is the company rather than the fruit, and that the surname in one document belongs to the same person as the surname in another.

That is one of three things the phrase is used to mean in AI-search writing, and separating them is most of the work of understanding this subject. The second is grounding in the retrieval sense — anchoring a generated answer to documents the system actually fetched, so the answer is not invented. That is what Google's Gemini API means by grounding with Google Search and what Microsoft means by grounding API results, and it has nothing to do with knowledge bases. The third is the marketed sense: publishing Organization schema, sameAs links and consistent naming in the hope of becoming a recognized entity and being favored by AI answers as a result.

The first two are documented mechanisms. The third is a tactic resting on the assumption that the first feeds AI citation. That assumption is the thing worth testing, and as of 29 August 2026 nobody has tested it.

Where the Knowledge Graph's contents come from

The Knowledge Graph is Google's database of things and the relationships between them; a knowledge panel is the box on a results page that displays one. Google announced it on 16 May 2012 at more than 500 million objects and more than 3.5 billion facts and relationships. Figures reported since — 570 million entities that December, 70 billion facts by mid-2016, roughly 500 billion facts on 5 billion entities in May 2020 — are second-hand, taken from Wikipedia's article rather than read from Google announcements, and Google has published no size figure since 2020.

What matters more than the size is the sourcing, which Google states directly. Facts in the Knowledge Graph "come from a variety of sources that compile factual information," plus licensed data for things like sports scores and stock prices. Knowledge panels "are created automatically by Google Search Algorithm when there is enough information available on the open web." There is no submission route. A verified representative can claim an existing panel and suggest changes to it; nobody can request that one be created.

That single documented sentence disposes of a whole product category. Anyone selling knowledge panel creation is selling work on the inputs — coverage, corroboration, consistent naming across sources — with no mechanism to compel the outcome, and the honest version of the offer says so. It is also worth keeping the graph and the panel apart. The panel is a display surface whose appearance varies by query, locale and confidence; its absence is not proof that no entity record exists, and its presence is not the record itself.

The one documented link to a generative surface

Google has connected the Knowledge Graph to an AI answer product exactly once. Announcing AI Mode on 5 March 2025, Robby Stein described it as drawing on Google's information systems, naming the Knowledge Graph, real-world information and the Shopping Graph's billions of product listings among its inputs.

Read what that establishes and what it does not. It establishes that the Knowledge Graph is among AI Mode's inputs. It does not say that entity membership affects which sources are selected, which pages are cited, or whether a brand is named in the generated text. And Google has never made the equivalent statement about AI Overviews: the May 2024 launch language describes core information quality systems without naming the Knowledge Graph, and no later source found as of 29 August 2026 fills that gap. Whether AI Overviews draws on the graph at all is unconfirmed.

Outside Google, the record is thinner still. OpenAI and Perplexity document nothing about entities. No operator anywhere states that Knowledge Graph membership, a knowledge panel, a Wikidata item or a Wikipedia article affects whether an AI answer cites or names you.

What Organization markup is documented to do

Google's Organization structured data documentation, last updated 15 April 2026, is precise about the function, and the precision is the useful part. Adding the markup to a home page "can help Google better understand your organization's administrative details and disambiguate your organization in search results." Some properties "are used behind the scenes to disambiguate your organization from other organizations," while others influence visual elements such as which logo appears. The same page carries an explicit disclaimer: "Google does not guarantee that features that consume structured data will show up in search results."

The documented function is disambiguation, not promotion. Those are different things and the difference is routinely erased in sales material. Google defines sameAs just as flatly — "the URL of a page on another website with additional information about your organization, if applicable." It is a corroboration pointer, here is the same thing described elsewhere. Google publishes nothing saying that the number or the choice of sameAs targets affects ranking, citation or panel eligibility. The practice is cheap and harmless; the mechanism attached to it in agency decks is invented.

For AI surfaces specifically, Google's AI features documentation says plainly, as of 10 December 2025, that "there's also no special schema.org structured data that you need to add." One controlled test agrees with it. Ahrefs tracked 1,885 pages that added JSON-LD between August 2025 and March 2026 against 4,000 matched control pages and reported on 11 May 2026 no meaningful uplift on any platform: AI Overviews −4.6%, AI Mode +2.4%, ChatGPT +2.2%. The limitation the authors flagged is usually dropped when the study is quoted — the sample was pages already being cited heavily, and all schema types were pooled, so the test says nothing about whether markup helps an invisible page become visible. Ahrefs sells SEO tooling and published a result that cuts against selling schema services, which is the circumstance in which vendor research is most credible.

The claim nobody has tested

Here is the load-bearing assumption of the entire entity-SEO-for-AI category: that being in the Knowledge Graph, or having a knowledge panel, measurably raises the chance of being cited or named in an AI answer.

Extensive searching on 29 August 2026 found no study measuring it. Not a weak study, not a contested study — none. What exists is agency and vendor writing asserting the connection, none of it presenting data. The claim is not false, either; nobody has shown that it fails. It is a hypothesis wearing the clothes of a finding, and it has been sold at scale for two years in that costume.

The design that would settle it already exists in the adjacent literature. Take entities that acquired a knowledge panel and match them against comparable entities that did not, then measure AI citation and mention rates before and after in both groups — the matched-cohort, difference-in-differences shape Ahrefs used to test schema. It is not an exotic study. It has simply not been run, and the absence is more informative than most of what is published about this subject.

Note what an absence of evidence costs a buyer. A tactic with no measured effect is not necessarily a waste; it is an unpriced bet. The problem is that it is usually sold as a priced one.

Wikipedia is cited, which is a different claim

The most common evidence offered for entity work is that Wikipedia gets cited heavily by AI surfaces, and the underlying observation is solid. Pew Research Center, in the only large non-vendor study of the question, found Wikipedia, YouTube and Reddit together supplying 15% of AI summary sources in its July 2025 measurement. Profound, across 680 million citations gathered between August 2024 and June 2025, put Wikipedia at 7.8% of all ChatGPT citations.

That is a finding about Wikipedia's own pages being used as sources. It is not a finding that having a Wikipedia article about your company causes an AI system to cite your site or speak your brand name. The two are presented as one claim constantly, and they are not related by any published mechanism. The tactic is also unavailable to most of the businesses being sold it, since Wikipedia's notability standards exist precisely to exclude organizations whose main qualification is wanting an article.

Source share of this kind is unstable in any case. Semrush's weekly tracking recorded ChatGPT cutting Wikipedia from roughly 55% of responses to under 20% in mid-September 2025, with no notice and no explanation. A strategy resting on which domains an operator currently favors is resting on something that changed inside a fortnight once already.

The terminology error behind most of the bad advice

The two senses of grounding get conflated more than any other pair of terms in this subject, and the conflation manufactures a conclusion nobody has stated.

When Gemini's API grounds a response, it runs searches and returns the queries it executed alongside citations tied to spans of generated text. When Microsoft refers to grounding API results, it means the same operation: generated text anchored to retrieved web documents. Microsoft's Bing Webmaster Guidelines, rewritten 26 February 2026, make eligibility for grounding results and citations a function of ordinary crawling, indexing and evaluation. Neither company documents knowledge-base membership as a gate on any of it.

Entity grounding, meanwhile, means resolving a name to a record. Slide the two words together and you produce the inference that being a recognized entity is a precondition for appearing in grounded answers — which no platform has said and no study has shown. It sounds technical, it uses both companies' own vocabulary, and it is wrong at the joint. This is the single most productive source of confident, unfounded advice in the subject.

What is actually worth doing

Two things in this subject are documented and cheap, and it is worth being clear that they are smaller than what is usually recommended.

The first is naming yourself unambiguously. Microsoft's guidelines of 26 February 2026 give the only direct instruction any platform has issued here: entity names should be clear and consistent, with no ambiguous references, alongside the companion rule that facts should be stated directly rather than implied. It is about your own pages, it costs an editorial pass, and it bites at the two places where identity plausibly matters mechanically — retrieval, where an inconsistent name makes it harder for any system to assemble the documents about you rather than your homonym, and generation, where a model composing an answer has to decide what to call you.

The second is Organization markup for what Google says it does: disambiguation and the features that consume it. Add it for that, not as a citation lever.

To check whether a machine-readable entity record exists for a name, the Knowledge Graph Search API returns matching entities in JSON-LD and is a reasonable instrument for the question. Two caveats travel with it. The API remains in Preview status as of 29 August 2026, under Google's pre-general-availability terms, and Google does not state that its coverage matches what Search's live systems use. A hit is evidence that a record exists. An absence is not proof that one does not.

Frequently asked questions