Reverse AI Search

Share of model: when ChatGPT cites you but Gemini doesn’t

“Cited by AI” is not one thing. The three big models retrieve differently and disagree constantly — so coverage has to be read per model, not in aggregate.

Updated May 20266 min read
The short answer

Share of model is your coverage broken out by individual AI — the share of a query set on which each of ChatGPT, Gemini and Grok cites you, read separately rather than blended into one number. It matters because the models disagree constantly: they pull from different retrieval sources, weight freshness and authority differently, and ground their answers in different ways, so a domain ChatGPT recommends can be entirely absent from Gemini. Aggregate “AI visibility” hides that — it can look healthy while you are invisible on a model your buyers actually use. Reading coverage per model exposes two things a blended score never will: which model is your weak flank, and how fragile any given win is. A query you hold on only one of three models is a single point of failure; a query held on all three is durable. The action is to widen single-model wins toward full coverage.

What is share of model?

Share of model is per-model coverage: instead of asking “does AI cite me?” it asks “does ChatGPT cite me? does Gemini? does Grok?” and keeps the three answers separate. It’s the answer-layer analogue of tracking your rank on different search engines rather than assuming they agree — except the disagreement between AI models is far larger than between search engines ever was.

For the short definition, see share of model in the glossary. This page is about why it matters and what to do with it.

Why do the models disagree so much?

Three reasons, broadly. First, retrieval: each model draws on a different mix of indexed and grounded sources, so the candidate set of domains differs before any ranking happens. Second, weighting: they value authority, freshness and structure differently, so even from the same candidates they pick different winners. Third, grounding: a model answering with live web grounding behaves differently from one answering from training alone. The net effect is that “cited by AI” is not a single fact — it’s three separate facts that often conflict.

What does per-model coverage look like?

The table below is an illustrative example for a single fictional domain — it is not measured data from any index. It shows the shape of the question: same domain, same query set, very different coverage per model.

Illustrative per-model coverage for one domain (example only — not measured data)
QueryChatGPTGeminiGrokRead as
queries like “best X for small teams”CitedCitedCitedDurable — held on all three.
queries like “X alternatives”CitedNot namedCitedGap on Gemini — your weak flank here.
queries like “X pricing”CitedNot namedNot namedFragile — single-model win, high-intent.
queries like “how does X compare to Y”Not namedCitedNot namedSingle-model win on Gemini only.

How should I read my own share of model?

Two patterns matter. A weak flank is a model where your coverage is consistently thin — if your buyers lean on that model, it’s your priority fix. A fragile win is a high-intent query held on only one model; losing that model loses the whole slot. The fragile-win pattern is especially important on pricing and alternatives questions, which you can identify using the intent taxonomy.

How do I widen coverage on a weak model?

Look at what the weak model cites instead of you on that query — the competitor domains and the source types it prefers. That is your evidence for what it weights. Then strengthen those signals on the relevant page. The diagnostic walkthrough is in why ChatGPT recommends your competitor and not you, and the rivals you appear beside on the models you do win are your competitor query overlap.

Why watch share of model over time?

Each model updates independently, so your per-model coverage drifts independently too. A win on all three can quietly become a win on two without any change on your side. Tracking each model separately on a cadence is how you catch a single-model lost citation before it spreads. For the broader data on how often the models diverge, see our State of AI Citations research.

See your coverage across all three models

Run the free Domain Check to read your coverage on ChatGPT, Gemini and Grok side by side — not blended into one number — and find your weak flank and your fragile wins.

Frequently asked questions

Why do ChatGPT, Gemini and Grok disagree on who to cite?
They retrieve from different sources, weight authority and freshness differently, and ground answers in different ways. The same question can therefore surface a different set of domains on each — disagreement is the normal state, not an error.
Isn’t one blended AI visibility number simpler?
Simpler, but misleading. A blended number can read “good” while you are absent on the one model a given audience uses most. Per-model coverage is the only way to see where you’re actually weak.
Is single-model coverage bad?
Not bad, but fragile. A query you win on only one of three models is a single point of failure — if that model shifts its sources you lose the slot entirely. Treat single-model wins as candidates to widen.
How do I improve share of model on a weak model?
Study what that model cites instead of you on the query — the competitors and the source types it favours — then strengthen the signals it weights. Start from why a model recommends your competitor and not you.