Entity Density for AI Citations: The 20.6% Benchmark and the Test That Broke It

Entity density is the most-quoted number in GEO that has never survived a controlled test. Kevin Indig's analysis of 1.2 million ChatGPT responses puts heavily cited content at 20.6% entity density against 5-8% for ordinary English text. Then the only experiment that actually varied density -- the GEO Lab's April 2026 test on Perplexity, 10 pages spanning 12.9 to 29.5 unique entities per 1,000 words, 22 pre-registered queries -- returned zero citations at every single level. Both findings are true. The reconciliation is the useful part: density multiplies recognition, it does not create it.

What entity density actually measures

An entity is a specific, identifiable thing with attributes and relationships. A keyword is a string. That distinction is the whole basis of the practice: Google drew the line publicly on May 16, 2012, when it launched the Knowledge Graph with 500 million objects and 3.5 billion facts under the slogan "things, not strings." Practitioners now estimate the graph holds more than 5 billion entities and 500 billion facts, though Google publishes no current count¹.

Entity density is simply the share of your words that are named things rather than filler. "Our platform helps teams work better" contains zero entities. "GeoAura tracks citations across ChatGPT, Perplexity, Google AI Overviews, Gemini and Claude" contains five. Measured mechanically, that is the gap between a page an engine can place in a knowledge neighbourhood and one it cannot.

The controlled test that returned nothing

This is the part almost every entity-density guide leaves out. In April 2026 the GEO Lab ran the cleanest test of the hypothesis available: 10 pages from a single domain, deliberately spanning 12.9 to 29.5 unique entities per 1,000 words, measured with a filtered spaCy named-entity pass that was locked before results were examined. Artur Ferreira pre-registered 22 competition queries where three to five of those pages counted as plausible answers.

Every query, every page, every density level returned zero citations. The correlation came back undefined for lack of variance. The preprint sits on Zenodo under DOI 10.5281/zenodo.19450361 with raw data on GitHub³.

"The same domain hit a 20% citation rate on a separate 30-query set — 6 citations, all in first position. Every one landed on a concept the site originated."

Read that second result carefully, because it is the actual finding. The domain did get cited — 20% of the time, all at position one — but only on concepts it originated. What earned those citations was not density. It was being the only source that could corroborate a specific claim. The same limitation applies to the foundational study everyone cites: Aggarwal et al. (KDD 2024) measured up to 41.5% visibility gains from adding citations, but their manipulations ran on pages already inside the retrieval candidate set. That bounds the result to domains that have already cleared recognition.

Same experiment, same week: structure worked, density did not

This is the comparison almost nobody runs, and it is sitting in the GEO Lab's own published protocol. Both variants used the same domain, the same Perplexity sonar-pro surface, the same 30 native-reference queries, and the same pre-registered scoring. The only thing that changed was the lever. The structural rewrite — declarative headings and answer-first opening sentences instead of narrative build-up — moved citation rate from 36.7% to 60%, a 23.3-point swing. The density manipulation — ten pages deliberately spread across a 2.3× range from 12.9 to 29.5 unique entities per 1,000 words, measured with a filtered spaCy en_core_web_lg pass that dropped generic cardinals, ordinals and time expressions — moved it from 0% to 0%.

Read the third row together with the first. Structure pays where the domain is already competing and pays nothing where it is not. That is not a contradiction; it is the shape of a gate. The GEO Lab's own interpretation names it: AI citation selection may run a two-stage test, first whether the domain is recognised enough to enter the candidate set, then which page is the best candidate. If stage one fails, optimising stage two cannot help — their analogy is polishing a CV that never reaches the hiring manager. The authors are explicit that this is a working heuristic, not a measured internal threshold, and that the result is specific to one domain, one query set and one Perplexity model³.

Chen & Koudas supply the complementary half at scale, using a design the GEO Lab could not have built from one site. By comparing pages of the same website, they remove domain reputation as a variable — every page starts from the same recognition base — and still find that formatting explains 2.8% of the variation between pages. Only one content feature survived their robustness tests: title and paragraph wording that matches the questions people actually ask. So the two studies agree from opposite directions. Entity density is a within-page variable, and within-page variables are competing for 2.8% of the problem. The 32% sits upstream, in whether the engine's own search returned you at all¹¹.

Why recognition gates density

Retrieval systems run an identity pass before a relevance pass. They match the query to entities, pull candidate passages, then check whether independent sources corroborate the claim before generating an answer. Machine Relations synthesized more than 680 million AI citations in May 2026 and described the mechanism plainly: AI engines select sources rather than rank pages, and the selection test is whether a brand appears across multiple independent contexts the engine can cross-reference⁴.

Peer-reviewed retrieval research supports the mechanic. Work on uncertainty-driven evidence selection shows retrieval-augmented systems favour evidence that reduces uncertainty, not evidence that maximizes relevance alone. An organization named in one source carries high uncertainty; the same organization named across five independent sources carries low uncertainty and wins the slot. ZipTie.dev's analysis found 96% of AI Overview citations come from sources clearing E-E-A-T credibility thresholds⁵. A fragmented brand identity raises assembly cost, and the engine cites the competitor it can verify faster.

Salience: why naming more entities is not the goal

The Google Natural Language API scores entity salience on a 0 to 1 scale — how central each entity is to the page's main topic. This is where most entity-density advice goes wrong. Stuffing thirty brand names into a paragraph raises density and lowers salience on all of them, because the model cannot tell what the page is about. Google's Enterprise Knowledge Graph formally supports only three entity types for linking — Organization, LocalBusiness and Person — so make those three unambiguous before chasing anything exotic⁷.

There is a measurable stylistic correlate. Citation-winning content is almost twice as likely (36.2% vs 20.2%) to contain definitive language such as "is defined as" or "refers to", which naturally introduces a named entity with clear attributes. Definition sentences do two jobs at once: they raise salience on the primary entity and they hand the model a liftable triple².

Where entity density does earn its keep

Density is not worthless; it is conditionally useful. Three conditions, each drawn from evidence rather than from a checklist.

1. 1.
   You are already in the candidate set for the query. Chen & Koudas found title and paragraph question-matching moves citation odds by roughly +14% to +15% among pages of the same site — a real effect, but an effect inside the retrieved set. Density behaves the same way: it differentiates among pages that are already competing.
2. 2.
   The entities you name are the ones the query is about. Naming OpenAI, Perplexity and Gemini on a page about ChatGPT citations adds three anchors that were already implied. Naming the specific standard, the specific study ID, and the specific dated release adds edges a model cannot infer. The GEO Lab's own 20% citation rate landed entirely on concept queries the site had originated — the entity being named was the site's own idea.
3. 3.
   The entity is stated in a sentence that survives being quoted alone. This is where the structural finding and the entity finding merge: the citation-winning form is a declarative sentence that names a specific thing and states what it does. "Entity density is the share of words on a page that are named entities" is that sentence. "Entity density matters a lot in today's AI landscape" is not, at any density.

How to raise entity density honestly: 6 steps

1. 1.
   Clear the recognition gate first. Before touching density, confirm the engine can resolve you at all: a Knowledge Panel, a Wikidata entry, consistent naming across LinkedIn and Crunchbase. If the identity pass fails, no amount of density matters. See how to get cited by AI search.
2. 2.
   Open with a definition sentence. "Entity density is the share of words that are named entities" beats any build-up. It raises salience on your primary entity and gives the model a liftable fact in the first 30% of the page, where 44.2% of LLM citations originate.
3. 3.
   Name the specific things, not the category. Write "OAI-SearchBot" not "the AI crawler"; "Gemini 3.8 Flash, deployed 2 September 2026" not "the latest model". Specificity is what turns a mention into an edge.
4. 4.
   State relationships as sentences. GraphRAG-style retrieval builds an entity graph from the corpus first. "Conductor analysed 167,867,680 AI Overview citations" contributes a real edge; an implied connection contributes an isolated node.
5. 5.
   Mirror it in sameAs, and let the prose stand alone. Point Organization and Person schema at Wikidata, LinkedIn and Crunchbase, but never rely on markup to carry a fact the prose does not state — models ground on rendered text, not on JSON-LD they never receive.
6. 6.
   Originate something. The GEO Lab's 20% citation rate landed entirely on concepts the site originated. Publishing a measurement nobody else has is the only entity signal that reliably survives the corroboration test.

How to measure your own entity density

"The web is a graph of entities and their relationships. Pages are just the surface through which those relationships are expressed."

What changed in September 2026

OpenAI released GPT-6 Astra on September 3, 2026 with multi-step web research built into the default reasoning path rather than exposed as an opt-in tool. Where earlier ChatGPT models relied on cached index snippets, Astra opens pages and reads them. That makes clean named-entity definitions a direct retrieval asset: a page that states what it is gives the model something to lift, and a page that relies on context fails the test⁸.

Two boundaries matter. First, 87% of ChatGPT Search citations match Bing's top-10 organic results, so classic index position still predicts eligibility — entity work raises your odds within the candidate pool, it does not create one. Second, ChatGPT mentions brands 3.2 times more often than it cites them with an attributed link, so entity clarity is doing more work for unlinked mentions than for clicks.

Frequently asked questions