When you wonder whether your brand shows up in AI answers, the first thing that comes to mind is to ask ChatGPT your own name. That tells you something, but very little: one answer, one moment. This article describes how we turned the same curiosity into a measurement that can be repeated every month: which questions we ask, what we record from each answer, how we write down the places we could not measure, and what data Google, Microsoft and Google Analytics give you today. The examples come from two measurements of our own panel: 23 September and 11 November 2026.
What does AI visibility measure?
Not one number but four separate questions.
Each answer comes from a different place, and none is enough on its own.
| Layer | What it tells you | Where it is measured | Its limit |
|---|---|---|---|
| Mention | Does your name appear in the text of the answer | A fixed question panel, by hand or with a tool | Each engine is asked separately; the same question can get a different answer each time |
| Citation | Is your page among the answer's sources | The same panel; Bing Webmaster Tools for Microsoft's surfaces | Your page can be cited while your name never appears in the answer |
| Impressions | How often your page was shown on Google's AI surfaces | The AI performance report in Search Console | Google's surfaces only; no ChatGPT or Perplexity |
| Visits and enquiries | Did the person the assistant sent visit your site, and did they write to you | The AI Assistant channel in GA4, plus your form records | A visit that carries no referring address can show up as "direct" traffic |
For the first two, the method you have is to ask the questions yourself. Search Console's report covers only Google's surfaces, Bing's covers Microsoft's surfaces and some partners it does not name; neither tells you whether your name appears in the text of an answer. The last two show whether being mentioned turns into a real visit and a message.
Why measure a rate rather than a rank?
Because AI answers have no stable order.
In research published in January 2026 by SparkToro and Gumshoe, an AI tracking start-up, 600 volunteers put 12 questions to ChatGPT, Claude and Google's AI answers a combined 2,961 times (November–December 2025). The result: the chance that ChatGPT or Google's AI gives the same list of brands twice for the same question is below one in a hundred, and the chance of the same list in the same order is about one in a thousand. The researchers themselves write that the work was not peer-reviewed, that volunteers used their own settings and that newer models may behave differently; they also note that the co-author works at Gumshoe, a company in the AI tracking field, and that this may be a conflict of interest.
Two practical consequences follow. First, a report that tells you "you rank third in AI" presents a snapshot of one moment as if it were a lasting position; the study's author openly criticises tools that report a rank. Second, the meaningful measure is how many answers you appear in: ask many questions several times each and write down the rate.
Our own panel falls short of that, and we are not hiding it: we ask each question once a month on the engines we can reach, and so far the only engine we have been able to reach is Perplexity. So we do not treat a change in a single cell as a trend. Our record also keeps where in the answer we appear, because appearing in the first sentence is not the same as appearing at the end of a list; but we do not report it as a lasting rank, and as the measure we use how many answers we appear in. If you can, ask each question several times and record the rate.
How do you build a question panel?
A panel is a fixed list of questions asked word for word every month.
Ours has 20 questions: 13 in Turkish, 7 in English. Four ask about our brand by name; sixteen never mention our name. We split the questions into four types, because each type measures something different:
- Brand questions (4): asked with your name; they measure whether the engine knows you correctly. For example, which services you offer or where you are.
- Service questions (8): asked without your name, the way a customer would ask; for example, asking for a local SEO agency in İzmir. This is where the real competition is.
- Sector questions (3): ask about the service together with an industry; they measure whether your work in that industry gets into the answer.
- Informational questions (5): ask about a topic related to your work; they measure whether your articles are cited as sources.
Each question has a target page: the page we expect to answer it. The question list has a version number; if we change a question, the version goes up and the comparison stops there. Six engines are defined: Perplexity, ChatGPT, Gemini, Copilot, Google's AI Overviews and AI Mode, and Claude. Twenty questions and six engines make 120 cells a month; a cell is one question's answer on one engine.
What do you record from each answer?
For each cell we write down:
- The date and the engine.
- The status: your name appeared, did not appear, or could not be measured.
- Place in the answer: where your name or your page appeared.
- Sources: the full list of links the answer showed, in order.
- Who was named instead of you: other firms that appear in the answer.
- Evidence: a screenshot or the text of the answer.
- A note: what the numbers mean; for example, which part of the answer a source supports.
Our own template taught us three lessons. First: in October, for non-brand answers, we recorded only the type of source (directory, company site, publication and so on; for eight of the twelve answers, with no type at all for the other four), not the full list; we started keeping source lists in November. So the two months' sources cannot be compared one to one. Second: in October we took no screenshots and saved no answer text to a file; the record holds only our notes, and today nobody can check those answers word for word. In November we saved the answer texts (for most, their first part) and the source links to files. Third: the template had no separate fields for "your name appeared" and "your page was cited". In November, in one informational question, our local SEO guide was first in the source list, but our name did not appear in the text of the answer; we had to separate the two with a note. If you are building a template from scratch, make them separate fields.
In November the one page of ours cited for a non-brand question was the Turkish version of our local SEO guide for small businesses in İzmir.Local SEO guide →Why doesn't an engine you could not reach count as zero?
Because zero is a result, and "not measured" is the absence of one.
Mix them up, and the month you finally reach that engine you will conclude that "our visibility went up" when it did not. In our own panel there are more unmeasured cells than measured ones:
- October (23 September 2026): 16 of 120 cells were measured. 80 could not be, because they required an account (ChatGPT, Gemini, Copilot, Claude); 20 because Google showed our browser a bot check; and 4 because Perplexity's logged-out query limit ran out.
- November (11 November 2026): 20 of 120 cells were measured. 80 again required an account, and 20 could not be measured because we did not try Google this month.
Our rule: we do not get around a login screen, a bot check or a sign-up layer covering the page. In November, from the second answer on, Perplexity covered the page with a sign-up layer; we did not remove it; we recorded the first part of the answer text sitting on the page underneath (in some cases all of it) and the source links. Because we have a screenshot of the first answer only, we do not quote that month's answers.
You can ask the engines that need an account from your own account. Then that account's history and personalisation settings may affect the answer; mark those cells separately and use the same account every month.
Get a free digital analysis for your brand — no sales pressure.
Free analysis →How do you compare two months?
By comparing only the cells measured under the same conditions in both months.
In November we measured 20 questions, in October 16; four English questions had hit Perplexity's query limit in October. Those four make November's non-brand total larger, but they have no October counterpart.
| Measure | October (23 September 2026) | November (11 November 2026) |
|---|---|---|
| Cells measured | 16 / 120 | 20 / 120 |
| Not measured: account required | 80 | 80 |
| Not measured: Google | 20 (bot check) | 20 (not tried) |
| Not measured: Perplexity query limit | 4 | 0 |
| Brand questions: our name appeared | 4 / 4 | 4 / 4 |
| Non-brand questions: our name appeared | 0 / 12 | 0 / 16 |
| Non-brand questions: our page was cited | Full source list not kept | 1 / 16 |
The right reading is this: on brand questions nothing changed, and on non-brand questions our name appears in no answer in either month. In one of the twelve non-brand questions measured in both months, the local SEO question, our guide was cited as a source in November; in October we had noted that our guide was not among the sources for that same question (we kept no source lists that month; the basis is the note in the record). Between the two measurements we had expanded that guide considerably, but a single cell cannot tell us that this is why it appeared. Because we ask each question once a month, this is an observation, not a trend.
The breakdown of both measurements by question type, and how the sources were distributed, is in our GEO guide.What is GEO? →What do Search Console and Bing show?
Google and Microsoft now give site owners data on their own AI surfaces.
These reports do not replace a panel, but they measure something a panel cannot: volume.
Search Console: the AI performance report
On 3 June 2026 Google announced separate performance reports in Search Console for its AI features; since 31 August 2026 they have been available to all sites. They cover AI Overviews, AI Mode and the AI features in Discover, and they break down by page, country, date and (for Search only) device. The only metric the announcement lists is impressions. This data was already part of the totals in the overall performance report; the new report shows it in a separate view. Google also writes that no extra markup, special file or special structured data is needed to appear on these surfaces.
Bing Webmaster Tools: AI Performance
In February 2026 Microsoft opened the AI Performance report in Bing Webmaster Tools as a public preview. It counts how often your pages are cited in Microsoft Copilot, in Bing's AI-generated summaries and in some partner integrations; it lists which of your pages are cited and a sample of the queries you were cited for. Microsoft states plainly that these numbers do not indicate ranking or importance, and the announcement includes no click data. Citation Share, added in June 2026, gives your share of all the citations shown for the same query.
The limits of the two reports: Google's covers only Google's surfaces; Bing's covers Microsoft's surfaces and some partners it does not name. Neither announcement mentions ChatGPT, Perplexity or Claude; answers on those engines still need a panel. Our own panel does not include these two reports.
How do you separate the visits AI sends you?
Google Analytics 4 has done this on its own since 13 May 2026: its default channel group has a separate channel called AI Assistant.
The channel looks at the address a visit came from; if it is one of the AI assistants Google recognises, the visit is assigned to this channel. Google's help pages name ChatGPT, Gemini, Claude, Copilot, DeepSeek and Grok as examples. No setup is needed.
It has two limits. Visits from Google's own AI Overviews and AI Mode are not assigned to this channel but to organic search. And a visit that carries no referring address, for example a link opened from an app, can show up as "direct" traffic.
Even so, a visit count is half an answer; the real question is whether the visitor wrote to you. On our own site the source of the first visit (the ad click ID, the UTM tags and, if the visit came from an AI assistant, that assistant's address) is stored in the visitor's browser and reaches us together with the message when a form is sent. We do not publish any figure from that record.
Five questions to ask before buying a measurement tool
There are tools that do this measurement for you, and they can collect far more samples than asking by hand. When choosing one, ask:
Get a free digital analysis for your brand — no sales pressure.
Free analysis →- 1.How many times do you ask each question? A one-off answer does not give you a rate.
- 2.Which route do you ask through: the interface users see, or the developer interface (API)? The two may not give the same answer; the SparkToro study also leaves this as an open question.
- 3.Which language and location do you ask from? A customer in İzmir asking in Turkish and someone asking in English from another country may not get the same answer.
- 4.How do you show an engine you could not measure? Are "not mentioned" and "not measured" kept apart?
- 5.Can I export the raw answers and the source lists? If all you see is a score, you cannot check how it was calculated.
If a tool reports your "rank" in AI, remember the research above: the same question rarely gives the same list.
What does a measurement month look like?
- 1.Start in the same week each month, with the same version of the question list.
- 2.Ask the questions in the same order on every engine. The one exception: put the questions that hit a query limit last month at the front this month.
- 3.Keep the evidence for each answer, a screenshot or the text, with its date.
- 4.Fill in the cells: status, the full source list, who was named instead of you, a note.
- 5.Take the month's numbers from the AI report in Search Console and from Bing's AI Performance report.
- 6.In your analytics, write down the visits in the AI Assistant channel and the messages those visits produced.
- 7.Compare only the cells measured in both months; wait a few months before tying a change to a cause.
Common mistakes
- Asking once and deciding.
- Writing zero in a cell you could not measure.
- Counting your name appearing and your page being cited on the same line.
- "Improving" the question list every month: when a question changes, the comparison ends there.
- Reporting a rank, or trusting a tool that reports one.
Measuring AI visibility is less about reaching a number than about keeping a record: the same questions, the same engines, the same fields, every month. Our record says something modest today: on the only engine we could measure, Perplexity, we are recognised correctly when our name is asked; when our service is asked, we are not there. We can write that because we measured it.
Let's measure your brand's place in AI answers with the same method.GEO & AEO service
