The Entity Play: How AI Decides You're Real

AI recommends, it doesn't rank

Start with the mechanic, because it changes everything downstream.

A search engine returns a ranked list. Ten blue links, ordered by relevance and authority, and the user chooses. The interface assumes the user will evaluate options. The engine's job is to surface candidates, not to decide.

An AI engine does something different. When a buyer asks ChatGPT for the best project management tool for a small agency, the model does not return a list of pages to evaluate. It returns a recommendation. It names two or three products, describes them, and often picks a favorite for the stated use case. The evaluation step that search delegated to the user has been absorbed into the model.

That absorption is the entire game. To get returned in a search result, you need a page that ranks. To get named in an AI answer, you need to exist in the model's understanding as a defined thing. A brand. A product. An entity with attributes, relationships, and a category. AI doesn't rank you. It recommends you, or it never mentions you at all.

The gap between those two outcomes is not gradual. In traditional search you can sit at position 8 and still catch traffic. In an AI answer there is no position 8. You are named or you are absent. Being seen and being selected used to be the same problem. Now they are two problems, and the second one is harder.

Entity recognition is the practice of getting an AI system to identify, categorize, and relate your brand to topics. It is foundational to generative engine optimization, because a model cannot recommend what it cannot resolve.

This is why entity optimization sits upstream of almost everything else in GEO. You can publish 200 articles. You can win the technical fundamentals. But if the model has no clean concept of who you are, those signals scatter. The mention is the signal, and the signal needs an entity to attach to.

What an entity actually is

An entity is not a webpage. It is a node in a graph.

AI systems build understanding from entities and the relationships between them, not from keyword matching. A model does not parse "best note-taking app" as a string to match. It resolves "note-taking app" as a category, retrieves the entities it associates with that category, weighs their attributes, and assembles an answer. Keywords are the query. Entities are the answer.

This distinction explains one of the most instructive findings in the field. Backlinko's entity SEO analysis looked at the query "note-taking tools." In traditional Google search, Evernote ranks number one. Its backlink profile and on-page SEO dominate the conventional result. But in AI recommendations for the same query, Microsoft OneNote consistently appears ahead. OneNote did not out-SEO Evernote. It out-related it.

OneNote sits inside the Microsoft ecosystem. It gets mentioned constantly in productivity discussions, enterprise software comparisons, and Office tutorials. Those mentions build dense entity relationships in the training data. The model has seen OneNote connected to "productivity," "enterprise," "Office 365," and "collaboration" thousands of times across thousands of independent contexts. When it reaches for a note-taking recommendation, OneNote is the entity with the richest, most reinforced web of associations.

"LLMs look beyond backlinks to evaluate context, co-occurrence, and repeated association between brands and topics." — Backlinko, 2025

That is the lesson in one sentence. Page-level authority wins rankings. Entity-level relationships win recommendations. The two strategies can produce opposite winners for the same query.

An entity has four things a model needs:

• Identity. Who or what is this? A defined, disambiguated thing with a name the model can resolve without confusion.
• Attributes. What are its properties? Category, founding date, headquarters, what it does, who it serves.
• Relationships. What is it connected to? Topics, competitors, use cases, partners, the contexts it shows up in.
• Disambiguation. How is it distinct from things with similar names? The signals that keep your brand from being confused with another.

Get those four right across enough surfaces and you stop being a string the model has to guess at. You become an entity it already knows.

Knowledge graphs are the foundation layer

Before a model can recommend you, something has to tell it you exist as a defined entity. That something is, increasingly, a knowledge graph.

Knowledge graphs are structured databases of entities and their relationships. Google Knowledge Graph, Wikidata, Wikipedia, and Bing's knowledge graph form the substrate AI systems lean on for entity recognition and disambiguation. They are not where most of your visibility comes from. They are where your identity gets established, so that everything else can attach to a defined thing.

Three layers matter most.

Wikidata is the open, structured base. It feeds directly into Google Knowledge Graph and influences multiple AI systems, and its notability bar is far lower than Wikipedia's. You can create a Wikidata entry with accurate properties, official website, founding date, headquarters, industry, key people, even when your brand does not yet qualify for a Wikipedia article. Update a fact in Wikidata and it propagates to Google in every language. For most companies, this is the single most accessible entity-building move available, and it is routinely skipped.

Wikipedia is the validation layer, and the numbers explain why it earns its own callout. Across the major studies, Wikipedia is the single most-cited domain in ChatGPT, appearing in roughly 27% of responses (Yext). A separate Profound analysis of 30 million citations from August 2024 to June 2025 confirmed ChatGPT sources information from Wikipedia more than any other site. A Semrush study put Wikipedia's citation frequency at 26.3% in mid-2025. The 5W Citation Source Audit found Wikipedia and Reddit together account for over 25% of all ChatGPT citations in the U.S., while The Wall Street Journal, The New York Times, and Bloomberg do not appear in the top 20. The legacy authority hierarchy and the AI citation hierarchy are not the same hierarchy.

The Wall Street Journal, The New York Times, Bloomberg, and the Financial Times do not appear in ChatGPT's top 20 cited domains. Wikipedia does, at the very top. — 5W Citation Source Audit, Q1 2026

A caution on Wikipedia, because the same studies show how unstable these surfaces are. Wikipedia's measured ChatGPT citation share swung from above 50% to under 20% after a September 2025 change, and Reddit's collapsed from roughly 60% to 10% in two weeks. The point is not the exact percentage on a given week. The point is that Wikipedia sits structurally at the center of how AI resolves entities, and that position has held across every methodology that has measured it. Note that you do not edit Wikipedia to promote yourself. That violates policy and can get you banned. You earn the notability that lets a neutral editor write the article, or you earn accurate mentions in the industry articles that already exist.

Schema.org is how you hand the graph a map of yourself. The load-bearing property is sameAs on your Organization schema, linking your brand to its Wikipedia page, Wikidata entry, LinkedIn company profile, Crunchbase listing, and verified social profiles. This is the mechanism that tells a machine "all of these scattered presences are one entity." Add @id to create a consistent identifier across your own site, knowsAbout to declare topic expertise, and Person schema for the executives and experts who carry your authority. Brands with managed Knowledge Panels see roughly 25% lower bounce rates, a downstream sign that clean entity data improves how the whole funnel resolves.

Consistency is the multiplier, contradiction is the tax

Here is the failure mode that quietly kills entity recognition. It is not absence. It is inconsistency.

AI systems cross-reference multiple sources to assemble an entity. When your brand name, leadership, product description, or category is stated one way on your site, another way on LinkedIn, and a third way on Crunchbase, the model does not average the three. It treats them as separate, weaker signals, or it lowers its confidence in the entity entirely. Any inconsistency in name, leadership details, or value proposition can cause AI to fragment what should be one reinforced entity into several thin ones.

The sameAs field exists to solve exactly this. It connects all your brand presences into one unified entity so the model stops guessing. But schema alone is not enough. The descriptions themselves have to match. The same one-line description of what you do, the same founder names, the same category language, repeated across every surface a model crawls. Repetition of consistent facts is what builds confidence. Contradiction is what erodes it.

This is also where consistency compounds into reach. The 129,000-domain study from Higglo and Digital Bloom found that brands present on four or more platforms posted 4.5x higher AI Share of Voice than brands confined to one or two. The same dataset found that active participation in relevant Reddit discussions correlated with 2.8x more AI citations than absence from Reddit, and that heavy discussion volume, measured in millions of mentions, correlated with 4x more citations than barely-mentioned brands. Breadth and depth of consistent presence are not vanity. They are the mechanism by which a model gains confidence that you are a real, established entity worth recommending.

The platforms that build the entity, each playing a distinct role:

• Wikidata and Wikipedia define the entity at the foundation.
• Google Business Profile supplies local and verification signals.
• LinkedIn carries professional authority and leadership data.
• Crunchbase provides structured business-entity facts.
• G2, Capterra, and Trustpilot deliver third-party validation.
• Reddit, YouTube, and X provide the active-conversation signal AI weighs heavily.

Four or more of these, all telling the same story, is the threshold where the multiplier kicks in.

The mention is the signal, and the source decides its weight

If knowledge graphs establish identity, mentions establish authority. And the most counterintuitive finding in the entire field is where AI engines look for those mentions.

Yext analyzed 6.8 million AI citations across ChatGPT, Gemini, and Perplexity. The headline result reframes the strategy: 86% of AI citations come from sources brands already control or heavily influence. First-party websites generated 44% of citations, business listings another 42%, and reviews and social a further 8%. This is not a license to ignore earned media. It is a correction to the assumption that AI visibility is something done entirely to you by third parties. A large share of the signal lives on surfaces you can directly manage, if the entity behind them is clean.

What you cannot control is which sources carry weight, and here the hierarchy is sharp. Onely's analysis of how ChatGPT chooses brands to recommend found a clear ranking of commercial signals:

• List mentions: 41%. Appearing in "best tool for X" roundups is the single strongest commercial signal. AI engines lean on curated lists as ready-made shortlists.
• Awards: 18%. Recognition from credible bodies feeds authoritativeness.
• Reviews: 16%. Aggregate sentiment on G2, Trustpilot, and the like shapes whether AI assesses you as recommendation-grade.

These are not arbitrary. They are formats that resemble the answer the model is trying to produce. A "best CRM for startups" list is a pre-assembled recommendation. A model reaching for that answer treats the list as evidence of category membership and standing.

The mention also has to be genuine, because the system is built to detect manufactured signal. You cannot fake authentic Reddit threads, invent podcast appearances, or game entity recognition the way PageRank was once gamed. The OneNote example proves the upside of the honest version. Its entity strength came from being mentioned in real productivity conversations across years, not from a campaign. The strategy is to become the kind of brand that gets mentioned, then make sure the entity those mentions attach to is unambiguous.

"Brand mentions have moved from a nice-to-have tactic to core infrastructure in an AI search environment. You can publish 200 articles, but without a citation footprint, LLMs have little reason to surface them." — Adam Tanguay, Search Engine Land

There is a sequencing point hidden in that quote. Mentions belong right after the technical and entity fundamentals are in place, before heavy content expansion. A citation footprint without a defined entity is noise. A defined entity without mentions is a node with no edges. You need both, in that order.

The entity-optimization sequence

Entity optimization is not a campaign. It is infrastructure you build once and maintain. The sequence below moves from the foundation outward, in the order that compounds.

1. Claim the graph. Create or correct your Wikidata entry first. Accurate instance of, official website, founding date, headquarters, industry, key people. This is the lowest-effort, highest-leverage move and it propagates.
2. Wire your schema. Implement Organization schema with a sameAs array linking Wikipedia, Wikidata, LinkedIn, Crunchbase, and verified social profiles. Add @id, knowsAbout, and Person schema for named experts. This is how you tell a machine your scattered presences are one entity.
3. Enforce consistency. Audit every surface for an identical brand description, founder names, and category language. Eliminate contradictions before you amplify anything. Inconsistency is a tax on every other signal you build.
4. Earn the foundation citation. Pursue Wikipedia notability through genuine third-party coverage, or earn accurate mentions in the industry articles that already exist. Never self-edit.
5. Build four-platform presence. Reach the 4+ platform threshold where Share of Voice multiplies. LinkedIn, Reddit, YouTube, and a review platform at minimum, each carrying the same entity story.
6. Earn the list mention. Get into the "best tool for X" roundups that drive 41% of commercial recommendations. This is where category membership gets ratified.
7. Speak in entities. Use specific brand names and proper nouns in your own content, not pronouns or generic terms. Give the model clean, extractable entity references to learn from.

Run that sequence and you stop hoping the model gets you right. You make yourself legible to it. The reward is not a ranking. It is recommendation, the only currency an AI answer pays out.

The brands that win the next phase of discovery will not be the ones with the most pages. They will be the ones the model can resolve without hesitation. Defined identity, consistent attributes, dense and genuine relationships, and a citation footprint that backs all of it up. Being seen was the old game. Being known well enough to be selected is the new one.

Every statistic in this piece follows Visigent's research standards: named sources, dated figures, corrections on request to help@visigent.com.