Skip to content
AI Visibility

How do I test whether my company is cited by AI?

Published on

To test whether your company is cited by AI, ask ChatGPT, Claude, and Gemini the questions your customers ask without naming you, several times each, and record who is cited and which sources appear. A single answer is not enough: when we asked the same questions again about an hour later, only 21 to 25% of the cited pages were cited again. This citability test is read as a percentage of presence, not as a position.

For what type of question? It all depends on the questions chosen for the test. A question that names your company measures what the AI finds on your site; a market question, which does not name you, measures whether the AI thinks of you. Both are useful, but they do not say the same thing, and a test that mixes the two is hard to read.

What those who measure recommend

The most widespread method comes down to three steps: define precise questions that a customer would ask without naming your company, ask them to several AIs, then note whether you are cited, which competitors appear, and which sources are cited. This is what AI visibility tracking tools do, in an automated way, for €20 excl. VAT to around a hundred dollars per month at the entry level.

Studies add a rule: repeat. In November and December 2025, SparkToro had volunteers ask the same recommendation requests to ChatGPT, Claude, and Google's AI answers: there is less than a one in a hundred chance that ChatGPT or Google's AI will give the same list of brands twice. Rand Fishkin concludes that an “AI rank” is meaningless, whereas a percentage of presence across many runs is a reasonable measure.

Choosing the test questions

The right questions are the ones your customers really ask, in their own words, at the moment they are looking for a solution: choosing a service provider, comparing two products, solving a problem in their line of work. A question that is too short or ambiguous is read in its best-known sense, and the cited sources follow that reading: on “Is what we read about GEO true?”, none of the three AIs first thought of optimization for AI engines: ChatGPT and Gemini thought of GEO magazine, Claude of organizations bearing this acronym. Each question must therefore be written the way a customer would ask it, with its context.

Next, separate two families. Questions that name your company, about your prices, your warranties, or your after-sales service, show what the AI finds on your site: on software troubleshooting, the documentation of the brand asked about accounts for 58% of ChatGPT's citations. Market questions, which do not name you, show whether the AI thinks of you: that is where most companies are absent, and it is the family that says the most about your visibility.

What to record, and how many times

For each answer, record whether your company is named, whether your site is cited as a source, which competitors appear, and which sources are cited. This last point is often the most useful: it shows where the AI goes looking in your market, and therefore where you would need to be talked about. With ChatGPT, about nine out of ten pages returned by the search are not cited; the cited sources are the ones that count.

Ask each question again several times, at different moments. On September 15, 2026, we asked questions to the three AIs, then asked them again identically about an hour later: none gave the same list of sources twice. The result of a test is therefore read as a frequency: your company appears in a given percentage of answers, for a given family of questions, with a given AI. A screenshot of an isolated answer says almost nothing.

The limits of a test done by hand

A test done in an AI's app can depend on the account used, on what the app remembers from previous exchanges, and on where the question is asked from; we have not measured this, and it is one more reason not to draw conclusions from a single person or a single session. The three AIs do not read the same web either: across our five combined corpora, only 1.8% of the cited domains were cited by all three. A test that covers only one AI says nothing about the other two.

Finally, a test measures a situation; it does not change it. If your company appears in no answer to market questions, repeating the test the following month will give the same result as long as nothing new has been published on the sources that AIs consult. The test serves to know where you stand and to choose what to do; if you already know that AIs do not cite you, one more measurement will change nothing.

What we do not know

Our own measurements cover the API of GPT-5, Claude Sonnet 4.6, and Gemini 3.5 Flash with web search. A test done in their consumer apps may give different results, and we have not measured it.

Sources

  1. SparkToro, Rand Fishkin — inconsistency of brand recommendations by AIs, .
  2. FaireDuBruit measurements, from September 4 to 16, 2026 — see “How we measured”.

Related questions

All questions about visibility in AI · Our offer