How to measure brand visibility in ChatGPT, Gemini, Perplexity, Claude and AI Overviews

A brand's visibility in AI answers is measured by counting individual responses: a question was asked to an engine, and the brand was either mentioned in the answer or it was not. This guide explains, in six steps, how to do that counting honestly. The method we describe is the one we use in our own measurements; we publish every weight and threshold on the How we measure page.

What are we measuring?

"Visibility" is not one thing. Your brand can appear in an AI answer in three different ways, and each should be measured separately:

• Mention: Does your brand's name appear in the text of the answer?
• Citation: Does the engine cite your site as a source when building the answer?
• How you are mentioned: As the leading option, as one of the alternatives, or with a reservation?

The first two are independent of each other, and this distinction matters more than people think. When we brought together the measurements of our own client and comparison brands on 3 August 2026, in 12.4% of the 2,748 responses that cited the brand's own site, the brand's name did not appear in the answer at all (in the first count we had found 2,733 responses and 12.8%; on 7 October 2026 we recounted the same responses). The engine may have read you without saying your name, and the customer reading the answer never hears of you.

Step 1: Write the questions without the brand name

This is the most basic rule of measurement. If you ask "X markası güvenilir mi?" ("Is brand X trustworthy?"), X's name will appear in the answer anyway; that is not evidence of visibility but an echo of the question. Visibility is measured by the questions a buyer asks before they know your brand, such as "İstanbul'da kurumsal müşterilere hizmet veren bir bordro yazılımı önerir misin?" ("Can you recommend a payroll software that serves corporate clients in Istanbul?").

Questions that contain the brand name are not useless: they show what the engine knows about you and how accurately. But they do not go into the visibility rate, and they need to be shown separately in the report. Combining the two in a single percentage makes the rate look higher than it is.

We applied this rule when measuring our own product too. On 27 August 2026 we asked five engines 26 Turkish-language questions in the GEO category; in none of the 116 measurable responses to questions asked without our product's name was our product mentioned. We published the result as it was on our What is GEO? page. Had questions with the brand name been added, this zero would have turned into a visibility that does not exist.

Step 2: Ask in Turkish, the way a buyer would

If buyers in Turkey ask AI in Turkish, the measurement should also be made with Turkish questions. The answer to a question in English does not represent the answer to a question in Turkish: the engine reads different pages and may mention different brands.

How the questions are written also shapes the result. In our method, questions are written the way five buyer profiles, researched for the brand's sector, would ask them, and they are divided into three types:

Once the question set is chosen, it is frozen. The next measurement is made with the same questions; otherwise the difference between two periods may come from the questions changing rather than from your work.

Step 3: Ask the five engines separately

ChatGPT, Gemini, Perplexity, Claude and Google AI Overviews do not give the same answer to the same question. One reason is that they look at the web in different ways when building the answer. According to OpenAI's help page, ChatGPT turns the user's question into one or more search queries and sends them to its search partners. Perplexity builds every answer with a real-time search and gives it with numbered sources. Google's AI Overviews runs inside Search and began rolling out gradually in Turkey from 18 February 2026.

Our own measurements show how large this difference can be. In the measurements of three corporate brands on 26 August 2026, the share of measured responses citing at least one source was, by engine:

This table is here not to rank the engines but to show why they cannot be combined. If half of one engine's responses cite sources while all of another's do, adding up and averaging the two engines' "citation rate" produces a meaningless number. That is why we report each engine separately and do not produce a single composite score.

Next step

Is your brand named in AI answers?

The visibility measurement asks five engines Turkish questions that do not name your brand, and records every answer. Answers that could not be measured are shown separately, never as zero.

See the visibility measurement

Step 4: Record every response and verify the mention

Every response should be recorded with its question, engine, date and full text. Keeping only the "mentioned / not mentioned" result means you will not be able to audit anything later. Behind every number you see in the report there should be responses you can click through and read.

Recording is not enough; you also need to verify that a mention really refers to you. A name appearing does not prove that the brand was mentioned:

• Namesakes: There may be another company, a person or a brand in another country with the same name. For example, the name "GoPeople" can refer both to an HR product and to a courier company in Australia.
• Echo: The answer may be repeating a word from the question.
• Position: Was the brand mentioned as the leading option, among the alternatives, or as an option to avoid? In our method a negative mention counts not as zero but as a minus.

A mention that cannot be verified is not counted in the brand's favour; it does not enter the calculation of the rate at all. In our method, if more than 20% of mentions could not be verified, no rate is given and "not measured" is written: a number calculated with that many gaps is an estimate, not a measurement.

Step 5: Do not count what could not be measured as zero

This step determines whether a measurement is honest. An engine may time out, a provider may return an error, a rate limit may cut a question off halfway. In that case you do not have the information "your brand was not mentioned"; you have no information. A report that turns the two into the same number claims a certainty that does not exist.

Every observation has three states, and all three should be written separately in the report:

Google AI Overviews is a good example. Google does not show an AI Overview for every search. If no AI Overview appeared for a search, that is the "measured, absent" state and is real information. If the service that fetches the search results returned an error, that is the "not measured" state and says nothing about your brand. If the two are written as the same zero, a service outage turns into a finding that "Google does not show you".

The numerical weight of this distinction is not small. In our measurement set of 26 August 2026, 189 (10.3%) of the 1,835 responses to questions asked without the brand name could not be measured or their source information could not be retrieved. Counting these as "did not cite a source" would have miscalculated every source's share.

Step 6: Repeat and read the interval

AI answers change even for the same question. In the research SparkToro published in January 2026, 600 volunteers asked 12 questions to ChatGPT, Claude and Google's AI a total of 2,961 times. In ChatGPT and in Google, the chance of the same list of brands coming back in two answers was below one percent. How often a brand appeared in the lists, however, was much more consistent than its ranking. The researchers also note that larger samples are needed.

Two practical conclusions follow from this:

1. Measure how often you appear, not your rank. The sentence "we are third on ChatGPT" may no longer hold for the next question; the sentence "we were mentioned in 12 of 40 questions" is sounder.
2. Repeat where there is uncertainty. In our method, questions on which the engines disagree are asked 3 more times to every engine that responded, and a majority decision is used. The result is given not as a single number but with an interval.

Example: how to read our own measurement

To see how the steps above turn into a report, let us look at our own product's measurement of 27 August 2026. The same figures are published on our What is GEO? page.

There are three things to note when reading this table. First, the result is not a single percentage but a pair of numbers: "0 of 116 responses". Second, the 9 observations that could not be measured sit on a separate row; had they been added to the zero, the report would also have counted 9 answers we did not measure as "not mentioned". Third, the zero is not a failing grade but a baseline: the next measurement will be made with the same 26 questions, and the change will be read against it.

Two other brands we measured on the same day with the same method gave different results: a digital signage product was mentioned in 6 of 97 responses (6.2%), and an HR assistant in 1 of 84. Because the three brands have different response counts, the results are written side by side but not turned into a single ranking.

Why a single score misleads

A single "AI visibility score" is appealing, but it usually hides three things: which engine it was measured on, how many responses it was calculated from and how many responses could not be measured. There is also no reliable Turkish data for setting the weights of the engines; a composite score calculated with weights that do not exist would be a made-up number.

That is why we write results as pairs of numbers ("0 of 116 responses"), show the engines separately and also say what a number does not show. A low visibility rate does not show:

• That consumer demand is low,
• That brand awareness or market share is low,
• That consumers do not prefer the brand,
• That AI does not recommend the brand (only that it did not mention it for these questions).

How to read the results

When reading a good visibility report, ask these questions in order:

1. Were the questions written without the brand name? Are the questions that include it shown separately?
2. Is each engine reported separately?
3. Does each rate state how many responses it was calculated from?
4. Were unmeasured responses counted separately, or mixed into the zero?
5. Was the previous measurement made with the same questions? If not, the two results cannot be compared.

If the answer to all five questions is "yes", the number in front of you is a measurement. If not, it is an estimate or a screenshot.

Frequently asked questions

How many questions are enough?

In our method, a default scan consists of 40 questions. If there are not at least 5 different questions asked without the brand name, no visibility rate is given at all: a rate calculated from a single question is not a rate. As the number of questions grows the interval narrows, but no question set represents a whole market; that is why the questions themselves are also listed in the report.

Can I run this measurement myself?

Yes. None of the steps above depends on a particular tool. The hard parts are not asking the questions but storing every response, weeding out namesakes and keeping unmeasured responses separate from zero. If you do this by hand, make sure you use the same questions, the same engines and the same rules in every measurement.