The State of AI Citations: What the Data Says About How AI Picks Its Sources

What is the State of AI Citations research?

Most commentary on AI search is anecdote dressed as data — a screenshot of one ChatGPT answer, a vendor’s single visibility number, a confident claim with no method behind it. This hub takes the opposite approach. Every finding here comes from the same place MentionRadar’s product does: a query–domain inverted index that records, query by query and model by model, which domains the major AI assistants cite or mention. Because the index is a measurement instrument rather than a one-off survey, it can answer questions a snapshot can’t — namely, what changes over time and where the three models diverge.

If you are new to the underlying mechanic, start with the Reverse AI Search pillar, which explains how an inverted index of AI answers is built and why it lets you look up a domain and read its AI citation footprint. This research hub is what happens when you point that same instrument at the whole index instead of a single domain.

Why measure AI citations at all?

Three of the most consequential questions in AI search are still answered with guesses:

• Do the models agree? If you optimise for ChatGPT, are you also winning Gemini and Grok — or are they citing entirely different sources for the same question?
• Is a win durable? If an AI cites you today, will it still cite you next week, or is the answer churning underneath you?
• What actually earns the citation? Is it backlinks, brand mentions, review-platform presence, content format — or some mix the hype articles never quantify?

You can’t make a confident content decision without an honest answer to these. The research below is our attempt to give defensible, method-first answers — and to flag, out loud, where the data is directional rather than conclusive.

The four things this index can measure that a score can’t

1. Model disagreement

Because the index queries ChatGPT, Gemini and Grok independently for the same question, it can compare their cited sources directly. The overlap (and the gap) is a metric almost no one publishes because almost no one queries all three at scale. See how often ChatGPT, Gemini & Grok disagree on sources.

2. Citation volatility

Re-running the same query over time shows how stable an AI answer’s source set is. Some answers are rock-steady; others swap citations week to week. That’s the difference between a defensible position and a coin flip — covered in AI citation volatility.

3. Content-type win rates

The domains and URLs the models cite fall into recognisable formats — comparison/“best of” lists, documentation, original research, community threads, review platforms. Which formats earn citations most often is an empirical question, addressed in which content types get cited most by AI.

4. Category share of voice

Within a category, a handful of domains tend to absorb most citations. Measuring that concentration gives a realistic benchmark for “good” visibility — see category share-of-voice benchmarks.

Our research principles (so you can trust the numbers)

• Method before number. Every finding states what was queried, how many models, and over what window. A number without a method is an opinion.
• Estimates are labelled. Where a figure is directional — a sample, a snapshot, a trend rather than a census — we say so explicitly.
• No fabricated proof. We don’t invent customer names, testimonials, or precise statistics we can’t stand behind. Claims are tied to the live index or to the three models’ observable behaviour.
• Contrarian when the data warrants it. If a popular belief — say, that prompt volume is a precise ranking input — isn’t supported, we’ll say so plainly.

When this hub starts publishing hard figures, every one will be sourced and dated. Our methodology explains exactly how a number gets from the index (or an attributed third-party study) onto the page — what is sampled, over what window, and how estimates are labelled.

The full research cluster

Every study in this pillar, in reading order:

• State of AI Citations 2026 (flagship report)
• How often do ChatGPT, Gemini & Grok disagree on sources?
• Which content types get cited most by AI (data study)
• AI citation volatility: how fast do AI answers change?
• Category benchmarks: average AI share of voice by industry
• The review-platform effect: G2/Trustpilot vs citation rate
• Does prompt volume mean anything? A data reality check

How to use this research

Read the findings as priorities, not promises. If the data says the three models disagree more than you assumed, optimise for all three rather than ChatGPT alone. If it says a category is highly concentrated, expect a longer climb and pick your queries carefully. The fastest way to make any of this concrete is to look at your own domain: the free Domain Check reads the same index backwards and returns the actual queries the three models already cite you on — your personal slice of the State of AI Citations.