AI Visibility: How I Measure It and With Which Tools

AI visibility is the share of AI assistants' answers to your customers' questions in which the brand is named, or its page is given as a source. I count it on my own question set, separately for each assistant and separately for mentions and citations. Nobody except the assistants' providers knows all the questions people ask them, so a “share of the AI market” cannot be calculated. Every number about AI visibility makes sense only together with what it was counted on: how many questions, in which assistant and on which day.

Three times between August and October my own check showed that a number that looked good meant something other than I thought:

1. The “brand presence” in a report from one of my projects turned out to be citations of the site, not mentions of the brand. In Perplexity the report showed 2%, while the brand was named in 37% of questions; in the questions without its name, in 21%.
2. A test that was supposed to show that position for sub-queries (queries that an assistant sends to a search engine itself while answering) predicts citations better than position for the main keyword gave a result consistent with the thesis, until a reviewer noticed that I had compared measurements of different accuracy.
3. A tool's citation count rose by half in a week for two brands, while their share barely changed, so most likely the base of answers collected by the tool (its index) was growing, not the brands' visibility.

I measure AI visibility so that such errors turn up before the client report, not after it. I described the second case in the post on AI SEO and GEO, and the other two are below.

What is AI visibility?

AI visibility is the presence of a brand in the answers generated by AI Overviews and AI Mode in Google, ChatGPT, Gemini, Perplexity and Copilot, counted as the share of such answers. A brand can be in such an answer in two ways, and I count them separately:

1. Mention: the brand name is in the text of the answer that the user reads.
2. Citation: the brand's page is given as a source, usually in a footnote or in a list of links next to the answer.

One does not follow from the other. An assistant can recommend a brand by name and not give its page, because it knows about it from memory or from other people's texts. It can also give the brand's page as a source of general information and never write its name. In the post on AI SEO and GEO I describe where an assistant gets its knowledge about a brand: from search or from the model's memory.

An AI answer has no position, so I measure a share, not a place. An assistant's answer is generated anew every time and is not a list of ten results. Eli Schwartz put it in one sentence:

So instead of a position I count in how many answers to my question set the brand appears. A number such as “the brand is in 3rd place in ChatGPT” describes one answer from one day, and the next answer to the same question may name other brands in another order.

How do you check your brand's AI visibility in ChatGPT and other assistants?

I check it in six steps and count only in the fifth. The earlier steps are there so that the number at the end concerns the right brand and the right questions.

Step 1. Check whether the assistant knows which brand you mean

Before I count anything, I ask a question with the brand name alone and read the answer. If the assistant confuses the brand with another company, every later number may concern the wrong company. For one Polish online store, an AI visibility tool showed 55 points out of 100 and a 100% share, and all the answers it was calculated from concerned another, well-known company with the same name. In another tool, a brand whose name is an ordinary dictionary word got over 400 thousand pages with a mention, because the tool counted every use of that word. The comparison brand had fewer than 7 thousand.

When a company from the same industry has the same name, the measurement loses its meaning, and when someone from another field does, it is only harder. In the first case, before I count anything, I put the brand's description in order on its site and in its profiles, so that the assistant can tell it apart. My own first and last name are shared by, among others, an actor and a pilot, so measuring my brand makes sense, but for every mention I check whether it is about me.

Step 2. Read the Google results for the brand name

AI Overviews and AI Mode use Google's index, so I first check what Google shows for the brand name alone. For every result in the first ten I write down whether I control that page fully, partly (for example a profile on someone else's site) or not at all. For my own first and last name on October 3, 2026 in Polish Google, I did not fully control any of the 10 results, and I controlled one partly.

Then I look in those pages for the sentences that assistants repeat about the brand. If a sentence is in the results, I know where the assistant gets it from and can ask for that page to be corrected. If it is not there, it is probably in the model's memory, and fixing pages will not change it quickly.

Step 3. Build a question set and do not change it

I take questions from demand data and from Search Console, not from a brainstorm. From Search Console I pick the queries the site already appears for, and I fill in the rest with keywords from keyword research tools. I described the split of questions by the site's position in the post on SEO strategy. When, in the period I take the queries from, a site has fewer than about 1,000 impressions and 100 different queries in Search Console, I do not derive questions from it, because from such a small sample I cannot pick questions that say anything about the market. It is a rule-of-thumb threshold from my process, not a test result.

I write most of the questions without the brand name, for the reason described in the post on AI SEO and GEO. The question “is X good” measures what the assistant knows about X, not whether it will recommend X to someone who has never heard of it. I write questions without the brand name in three forms (the examples are illustrative):

• a category with a qualifier, for example “accounting firm for a one-person business in Kraków”;
• a comparison that contains only a competitor's name, for example “alternative to [competitor]”;
• a question about a specific capability, for example “which note-taking app works offline”.

I count reputation questions separately and do not include them in visibility. “Is X a reliable company” and the pairs of questions “who to choose” and “who to avoid” show what the assistant thinks of the brand, not whether it recommends it. For the store I wrote about in step 1, the assistants answered the question of whether it is a reliable store by citing customer complaints.

I ask questions about a Polish brand in Polish and in English. In my test some of the sub-queries for Polish questions were in English, which I describe in the post on AI SEO and GEO.

I do not change the set between measurements. I add new questions as a separate group, because every change to the list breaks the comparison with the previous measurement.

Step 4. Set the number of repetitions with a pilot and record when and where

I do not set the number of repetitions in advance: on a new question set I first run a pilot in which I ask each question 10 times, all at the same time. This is what Dmitrij Żatuchin recommends in the Dice Roll Method protocol for repeatedly querying models about brands (arXiv, version 2 of October 14, 2026). From the pilot you calculate how much the answers to the same question differ and how much the questions differ from each other, and from that how many repetitions you need for the result you care about. In this version of the paper the author withdrew the fixed thresholds from version one (from 5 to 20 repetitions depending on the type of study), because it turned out they had been calculated on data from a synthetic generator, not on data from studies.

A pilot usually gives two different numbers: few repetitions for a result for the whole assistant, and many more for comparing a single question before and after a change. I show the numbers from my test and from Żatuchin's paper below, in the section on repetitions. The AI Overview was the most stable in my data: for 77 of 94 questions it gave three identical answers, so one pull is usually enough for it.

For every answer I record four things: the assistant, the surface, the model version, and the date with the time. The surface is the place where the model answers: a chat app, the API, AI Mode in Google or an AI Overview. Answers from the API need not be the same as what a user of the app sees, so I do not add results from two surfaces into one number.

The time of an answer lets me tell randomness from a change over time. In the repetition test Perplexity cited the project's site in 0, 2 and 9 questions in successive repetitions, with a nearly constant number of mentions (33, 36 and 35). Most likely access to the site changed during the measurement, but the spreadsheet had no answer times, so it cannot be checked. That is why in later measurements I send the repetitions at once and record the time of every answer.

Step 5. Count mentions and citations separately

The result for an assistant is the share of all answers in which the brand appeared, separately for mentions and for citations. I add up the sources from all repetitions. Questions where the brand appeared in only some of the answers I mark as unstable and draw no conclusions from them about a single question.

Step 6. Check what could invalidate the diagnosis

Before I tell a client what they are missing, I check the situations in which the diagnosis would be wrong:

• The assistant recommends nobody in this category, for example for questions about health, finance or law. Then there is no gap the brand could fill.
• The assistant answers from memory, without searching. Then work on the site will change nothing for a long time.
• A competitor that I put into the comparison myself does not work. In one measurement, four domains on which I based the thesis that the category was empty did not even have a DNS record, that is, the domain address led to no server. Zero citations for a competitor whose site does not work do not mean a free place in the category.
• AI visibility is rising while Google traffic in the same period is falling. SearchPilot describes a test for the site Omio in which a change on the pages helped traffic from language models but, by their estimate, would have lowered Google traffic by about 6.5%, so it was not rolled out (SearchPilot). That is why I always look at Google traffic next to AI answers.

After every change on the site or in the brand description, I repeat steps 4 to 6 on the same question set. Only then can the result before and after the change be compared.

Which AI visibility metrics should you measure?

For each assistant I measure two numbers: the share of answers with a mention of the brand and the share of answers with a citation of its page. On top of that, the share against competitors on the same question set and the number of unstable questions. I treat traffic from assistants in analytics as a separate measure, because it is always understated: some visits arrive without source information.

When I counted mentions separately, it turned out that the “brand presence” report had been measuring citations. In one of the projects I run, the measurement of August 29, 2026 covered 94 questions, five AI surfaces and three repetitions, 1,410 answers in all. The script that calculated it counted a brand as present only when its page was among the sources, while the description of the method in the report spoke of the brand's presence in the answer. In October I counted the brand name alone in the same answers:

In Perplexity and Gemini the brand was named in more than a third of the questions, while the report showed 2% and 0%. The mention figures include 19 questions with the brand name, so in just the 75 questions without the name they are lower: from 21% to 28% depending on the assistant. That is still roughly one question in five in assistants where the report showed almost zero. Since this analysis I count both measures separately and demand the same of every report I receive.

A mention without a citation and a citation without a mention are, for me, two different hypotheses and two different jobs, which I describe in the section on the visibility gap. I test both hypotheses with a change and the next measurement.

Share counted per answer, not per citation

I count a brand's share among competitors per answer, not per citation, because answers have very different numbers of sources. In the measurement for one Polish SEO agency, 23 answers to questions about recommending an agency had from 1 to 13 sources. A share counted per citation would give an answer with 13 sources thirteen times the weight of an answer with one. Counting per answer, every answer weighs the same, because a user reads one answer at a time.

The source list from ChatGPT has to be cleaned of search results that the model only looked at. The tool I use to collect ChatGPT answers returns, next to the sources cited in the answer, the whole list of search results that the model saw. In one answer there were 52 such results and 7 cited sources. In the same 23 answers, after discarding that list, the number of citations fell from 279 to 181, and the list of places worth appearing in from 10 to 6 entries.

How many repetitions are enough?

In my test three repetitions of each question were enough for a result for the whole assistant, but not for assessing a single question. In the repetition test about half of the questions in which the brand appeared at least once had an unstable result: in ChatGPT 18 of 36 for the citation of the site (all 94 questions), and in AI Mode 16 of 29 for a mention in questions without the brand name. A question that “has the brand” in one run may not have it in the next.

• 13%The share of questions without the brand name that would be wrongly classified in ChatGPT (the brand is named or not) with three repetitions. When asked once, 21%, with five 9%, with seven 8%.Model estimate on data from 1,410 answers, October 14, 2026
• 1.5 percentage pointsThe typical error (standard deviation) of the share of answers with a brand mention in ChatGPT with three repetitions, in questions without the brand name. With five, 1.2 percentage points.Model estimate on data from 1,410 answers, October 14, 2026
• 0.33The median similarity of the lists of cited domains between two repetitions of the same question in ChatGPT and AI Mode: the overlap divided by all domains from both lists (1 means identical lists). In Perplexity 0.92.Repetition test, October 14, 2026
• 77 of 94The number of questions for which AI Overview gave three identical answers.Repetition test, October 14, 2026

Going from three to five repetitions improved the result for an assistant in my test by less than a percentage point, so for this result I usually stay with three. It is an estimate from a statistical model fitted to one project, one measurement window and two assistants (ChatGPT and AI Mode), not a measurement, and that is why on a new set I check it with a pilot. Żatuchin has a similar result: on 250 different questions, one or two repetitions already gave a reliability coefficient of 0.80, and five repetitions 0.96 (1 means that the error-free means of the repetitions tell the questions apart).

To compare a single question before and after a change on the site you need 15-25 repetitions, not three. According to Żatuchin's simulation, with five repetitions the chance of detecting a large change on a single question is only 15-19%, and 80% is reached only at 15-25 repetitions. His data concern models queried through the API about recommending brands, so for answers with search I check that number with a pilot too.

I do not use the source list from a single repetition, because in ChatGPT and AI Mode the overlap of two repetitions of the same question is usually only a third of all the domains cited in both. I build the list of places where the brand should appear from all repetitions, with the number of appearances of each domain.

Sources can also change from week to week. SISTRIX tracked 82,619 questions for 17 weeks in six countries (not including Poland), and in ChatGPT up to 74% of cited sources changed from one week to the next; that is the highest value, not the average (SISTRIX, May 1, 2026). That is why I accept a change in sources only when it repeats in the next measurement.

How do AI visibility tools work, and how do you choose one?

AI visibility tools do the same as a manual measurement: they ask the assistants a set of questions, record the answers and count where the brand appears. They differ in whose questions they are, how many times they ask them and what they count. I take data from three kinds of sources:

Absolute numbers grow together with the tool's index

A tool's citation count can rise by half in a week even though a brand's share does not change. In an Ahrefs report on the number of AI answers that cite a domain, based on Ahrefs' question base, I compared two readings for two large online stores in Germany, from July 26 and August 2, 2026. The number of their citations in ChatGPT rose by 50% for both in that time, while one store's share of the two stores' combined citations barely changed: 34.2% and 34.1%. In Perplexity the numbers rose by 68% and 185%. Most likely the tool's base was growing, not those brands' visibility. It is two brands, one tool and one week.

That is why from tools I take the share, not absolute numbers, and only for assistants where the share in my comparison did not change by more than half a percentage point. That was the case in ChatGPT, AI Overview, AI Mode and Grok (xAI's assistant), while in Perplexity, Gemini and Copilot the share changed by about 3 to 11 percentage points. It is one comparison from one week, so I assume stability until the next reading contradicts it.

What I check before choosing a tool

Before choosing a tool I check seven things:

1. Whether it shows mentions and citations separately.
2. Whether I can add my own questions and keep them unchanged for many months.
3. How many times it asks each question and whether it shows which questions are unstable.
4. Whether it records the model version, the surface and the date of every answer.
5. Whether it supports my market and language.
6. Whether it shows full answers, so that I can check that it is about the right brand.
7. Whether it gives the share on the same question set, not only absolute numbers.

I do not report sentiment scores from the classifier I tested. In one case it marked as negative an answer that ended with the verdict “X is the better choice for most teams”. I assess sentiment by reading the fragment of the answer itself, or I write nothing about it.

A single “AI visibility score” from a tool is not my goal. It combines mentions, citations and different assistants, so you cannot read from it what changed. I always give a number from a tool together with a business measure: inquiries from customers, sales or Google traffic.

Free sources: Bing and Google

Bing Webmaster Tools shows for free how many times Copilot and AI answers in Bing cited your pages. The AI Performance report was announced on February 10, 2026 as a public preview. It shows the number of citations, the number of cited pages per day, citations of individual URLs and a sample of the phrases the AI used to look for content before it cited it (Bing Webmaster Blog). This is data from Bing and Copilot only, so it does not replace a measurement in ChatGPT.

Search Console shows impressions of pages in AI Overviews and AI Mode, but it does not say whether the brand was named. Google made this report available to some sites in June 2026, and to everyone from August 31, 2026 (Google Search Central, Search Console Help). The description of the report speaks of impressions by page, country, date and device, and does not describe mentions in the text of an answer or the questions that triggered them. It is a good check on Google's side, but not a measurement of mentions.

Your own measurement through the API

Your own measurement through the API costs little, but it needs time to set up and check. A run of 14 questions on four surfaces (ChatGPT with search, AI Mode in Google, and Gemini and Perplexity through the API), 56 answers in all, cost me USD 0.82 in API fees, of which USD 0.63 was Gemini (I gave the first of these numbers in the post on AI SEO and GEO). It was the first of three measurement runs for an SEO agency, and the whole measurement, 180 answers in all, cost about USD 2.8. At those rates a set of 100 questions on four surfaces with three repetitions, that is 1,200 answers, would cost about USD 18-19 per measurement, mostly for Gemini. A pilot with ten repetitions on the same set is 4,000 answers and about USD 60, one time only. What takes time is writing the questions, checking that the assistant does not confuse the brand with another, and reading the answers. In the post on AI tools for SEO I described the tests after which I calculate numbers from such measurements with a script, not with a model.

What do I do when the measurement shows a visibility gap?

The measurement shows what situation the brand is in, and what I do depends on that:

• The brand is neither in mentions nor in sources, and the assistant searches the web: I work on the site's visibility in Google for those questions and their variants. I start by checking whether the assistants' bots get the site at all, as in an SEO audit.
• The brand is named, but its page is not cited: I fix the fragments of the page so that they answer the question without the rest of the text.
• The page is cited, but the brand is not named: I work on what others write about the brand, because it is probably from other people's texts that the assistant knows whom to recommend. In the post on link building in the AI era I described how I plan mentions on sites that AI already cites.
• The assistant answers from memory and gives outdated or untrue facts: first one consistent description of the brand in its own channels, then publications in places edited by others. I do not promise an effect within a quarter.

I check the effect of each of these jobs with the next measurement on the same question set, because none of them comes with a guarantee. In the post on AI SEO and GEO I described which of these actions have evidence and which are popular advice with no measurement. My tests that did not confirm several such pieces of advice are there too.

What won't AI visibility measurement tell you?

The measurement will not tell you how many people saw the brand in AI answers. The share concerns my question set, not all conversations with assistants, whose number I do not know.

It will not tell you why the assistant chose one brand and not another. A diagnosis from the measurement is a hypothesis that I test with a change and the next measurement on the same set.

A result from one day is not a trend. Providers change models and the way they search, and the changes are not always announced, so I compare only measurements with the same surface and a recorded model version, and I accept a change only when it repeats in the next measurement.

For Poland there is no data on which assistants people use. I did not find a Polish study of how often people use each assistant, so I do not know how much weight a result in each of them carries. That is why I show the result for each assistant separately and do not combine them into an average.

What do I measure on my own brand?

On rogala.pro I measure the same as for clients, but the first measurement (the baseline) of July 26, 2026 I did with a weaker method. It had one answer per question and did not include results from ChatGPT. I described the baseline and what I have changed since in the post on AI SEO and GEO. On August 25 I set up daily checking in Ahrefs Brand Radar of the four baseline questions in ChatGPT for Poland. I do not describe the results of that daily checking until I put them next to the measurement of November 1. I planned the first measurement with the method from this post for November 1, 2026 as a pilot: the same four questions on four surfaces, each 10 times. I will describe the result in both posts, even if it is zero.

Frequently asked questions about AI visibility

Where to start?

You can do a first measurement yourself in an hour, without a paid tool. Ask an assistant one question with the brand name alone and check whether it knows who you mean. Pick ten queries from Search Console that your site appears for, ask them without the brand name three times in ChatGPT and in Google's AI Mode, and write down separately where the brand is named and where its page is the source. That gives you a picture of the whole set; a pilot with ten repetitions will tell you more about single questions. If the result surprises you, it is a good topic for a consultation.