Does Gemini Decide What Goes Into Google's Knowledge Graph?

Two systems, a decade apart

The question assumes a relationship that the timeline rules out. Google announced the Knowledge Graph on 16 May 2012, describing it as a model that “understands real-world entities and their relationships” and stating that it already held more than 500 million objects and more than 3.5 billion facts. Gemini arrived more than a decade later. Whatever built the graph, it was not Gemini.

The two things are also different kinds of object. The Knowledge Graph is a structured database: entities, properties, and relationships between them, queried and displayed. A large language model is a generative system that produces text. One stores facts in fields; the other predicts language. Asking whether Gemini chooses Knowledge Graph entries is a little like asking whether a spellchecker chooses what goes in a dictionary.

This is not pedantry, because the practical advice that follows from each assumption is different. If a generative model curated the graph, it would be reasonable to try to influence the model. If the graph is a database fed by sources Google does not fully disclose, the only honest advice concerns being a clearly described, consistently identified entity — which is what Google's documentation actually addresses.

It is worth being precise about what the graph is claimed to contain, because the 2012 figures describe objects and relationships rather than documents. Google said the system held “more than 500 million objects” and “more than 3.5 billion facts about and relationships between these different objects”. Those are database rows about entities, and the entities in question were largely drawn from existing public reference material rather than generated from scratch by any model.

What Gemini's grounding documentation actually describes

Google publishes a detailed account of how Gemini obtains external facts, in its Grounding with Google Search documentation. It says the feature “connects the Gemini model to real-time web content”, so applications can “increase factual accuracy”, “access real-time information” and “provide citations”.

The described workflow has three steps. The model analyses the prompt to decide whether a search would improve the answer. Then “the model automatically generates one or multiple search queries and executes them”. Then “the model processes the search results, synthesizes the information, and formulates a response”.

The returned object is equally specific: search metadata recording the queries run, citations delivered as “inline annotations directly on the text content block” with start and end indices identifying which span each citation supports, and a search entry point for rendering suggestions in the interface.

What that documentation never does is mention the Knowledge Graph. Not as an input, not as an output, not as a system Gemini consults, edits or populates. The grounding path Google documents runs through Search and the live web. If Gemini were the mechanism by which entities enter the graph, this is the document in which that would appear, and it does not.

What Google tells site owners about AI Overviews and AI Mode

Google maintains a page specifically for site owners asking how to appear in AI features. Its AI features documentation is unusually direct: there are “no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary”.

It goes further and closes off the most commonly sold tactics by name: “You don't need to create new machine readable files, AI text files, or markup to appear in these features.” That sentence is aimed squarely at a market selling AI-specific files and schema as the route into generative answers.

The controls Google does document are all about exclusion rather than promotion: robots.txt directives for Googlebot, the preview controls nosnippet, data-nosnippet and max-snippet, noindex, and Google-Extended, which limits AI training and grounding in some of Google's other systems. The documented levers turn visibility off, not on.

This page also makes no mention of the Knowledge Graph as a factor in what appears in AI Overviews, and does not describe structured data as an inclusion criterion beyond general search practice. Two separate Google documents about its generative systems both decline to connect those systems to the graph.

Where the confusion comes from

The conflation is understandable, because the surfaces sit next to each other. A search for a company can return an AI Overview and a knowledge panel on the same screen, both presenting summarised facts in a box. From the outside they look like one system with one source.

Vocabulary makes it worse. “Knowledge graph” has become a generic term in marketing for any structured representation of entities, including internal ones used by AI vendors. Google's Knowledge Graph is a specific proper noun. A claim that a model uses a knowledge graph is not a claim about Google's Knowledge Graph, and the two get elided constantly.

There is also a commercial incentive to elide them. A service that cannot promise a knowledge panel can promise to make you “legible to Gemini”, and if the two are treated as one system, the same deliverable appears to satisfy both. Keeping them separate makes it obvious that neither outcome is documented as purchasable.

The deprecation of the Knowledge Graph Search API cuts against the rebuilt-by-AI story as well. Google is migrating that interface to a Cloud Enterprise product and warns the old one “is not suitable for use as a production-critical service”. That is a system being wound down and moved, not one being reconstructed by a flagship model.

Does the Knowledge Graph still matter, then?

Yes, but for what it has always done. Panels still appear, and Google still defines them as boxes shown for entities “that are in the Knowledge Graph”. Entity disambiguation still determines whether Google treats two mentions of a name as the same organisation or two different ones, which matters regardless of which surface is displaying the answer.

What Google documents as being within your control is the same set of things it documented before generative answers existed: describe the organisation accurately on your own site, use Organization structured data including sameAs URLs to profiles you genuinely operate, and keep identifiers and naming consistent wherever they appear. Google attaches an explicit no-guarantee clause to all of it.

The useful mental model is that generative answers and the Knowledge Graph both benefit from the same underlying condition — an entity that is consistently and unambiguously described across the open web — without either one being a lever the other operates. That condition is slow to build and cannot be purchased as a discrete deliverable.

It is also worth separating two questions that get merged in practice. Whether an entity exists in Google's graph is one question; whether a given generative answer mentions your company is a different one, resolved at query time by whatever the system retrieves and synthesises. A company can be perfectly well described in structured terms and still go unmentioned in an answer, and a company with no panel at all can be quoted in one, because the two surfaces are populated by different processes that Google documents separately.

None of that makes entity hygiene pointless. It makes the payoff indirect and unguaranteed, which is a materially different sales proposition from the one usually offered. The work is cheap, durable and within your control, and Google describes it as helping the system disambiguate you. It is simply not a purchase order for a panel or for a mention in an AI answer.

What this means for anyone buying coverage

If a vendor tells you that Gemini selects Knowledge Graph entries, they are describing a system that does not exist as described, and any method built on that description is untestable. Ask which Google document says it. There isn't one, and the two most relevant documents both omit the graph entirely.

Be equally wary of the inverse pitch: that specific markup, an AI-specific text file, or a particular content format will secure inclusion in AI Overviews. Google has written the opposite in plain language, stating that no additional requirements or special optimisations are necessary and that new machine-readable files are not needed.

Media Morpheus sells media and podcast placement, and this finding is inconvenient for the category we operate in. The documented controls Google publishes for its AI features are mechanisms for opting out. There is no published mechanism for opting in beyond being ordinary, indexable, accurate content — and we would rather say that than sell a story about a model that Google's own documentation contradicts.

Methodology and limitations

This report compares public Google documentation as of 23 September 2026: the Knowledge Graph launch announcement of 16 May 2012, the Gemini API Grounding with Google Search documentation, the Search Central guidance for site owners on AI Overviews and AI Mode, and the deprecation notice on the Knowledge Graph Search API. The core claim — that Google does not describe Gemini as building or selecting Knowledge Graph entries — rests on absence across the two documents where such a description would belong. Absence in documentation is not proof about internal systems; Google does not publish how entities enter the graph, and it is entirely possible that machine learning of some kind is involved in maintaining it. What we can establish is that no published Google document supports the claim, and that the timeline makes the strong version of it impossible. Every URL cited was retrieved successfully before publication.

Conclusion

The Knowledge Graph predates Gemini by more than a decade, is a structured entity database rather than a generative model, and is not mentioned anywhere in Google's documentation of how Gemini grounds its answers. Google's guidance to site owners on AI Overviews and AI Mode states that no additional requirements or special optimisations are necessary, and that new machine-readable files and markup are not needed to appear. Meanwhile Google's public Knowledge Graph API is deprecated. The premise that a generative model curates the graph is not supported by any of this, and advice built on it cannot be tested, delivered or audited.

Frequently Asked Questions

Sources

1. official Google AI for Developers. Grounding with Google Search for the Gemini API: the query-generation workflow, inline citation annotations and search entry point. Contains no reference to the Knowledge Graph.
2. official Google Search Central. Guidance on AI Overviews and AI Mode for site owners, including the statements that no special optimisations are necessary and that new machine-readable files are not required.
3. official Google, 16 May 2012. Introducing the Knowledge Graph: things, not strings. Establishes the launch date and original scale of the Knowledge Graph.
4. official Google for Developers. Knowledge Graph Search API deprecation notice and migration to Cloud Enterprise Knowledge Graph.
5. official Google Search Help. Knowledge panel definition, automatic generation and the general categories of sources that inform panel content.
6. official Google, 20 May 2020. Danny Sullivan, A reintroduction to our Knowledge Graph and knowledge panels: over 500 billion facts about five billion entities. Read October 10, 2026.