AI Tracking: How to Track Brand Mentions in AI Search

Yes, by sampling. There is no Search Console for ChatGPT, and no AI company reports how often it recommends a brand. What you can do is ask the questions yourself, on a schedule, and keep count. That gives you a sample rather than a census. A few hundred answers is enough to see who gets named and who does not. Four things are worth recording for every answer: whether the text names your brand, whether it links to your site as a source, where in the answer you appear, and which other brands it names. The first two are separate events. An answer can link your page and recommend a competitor in the same paragraph, which is why we treat them as two metrics (cited isn't named). The term has other meanings too, such as following objects in video or monitoring how staff use AI tools. This article is about brands in AI answers only.

Because the answer is often the shortlist. Ask ChatGPT for the best tool in a category and it names three to six, with a line on each. A buyer who reads that list may never search for you by name. Your analytics will not show it. GA4 records a visit only when someone clicks a link in an answer, and most answers end without a click. Search Console now reports impressions in Google's AI Overviews and AI Mode, but only for Google's own surfaces, and an impression is not a recommendation. If you want to know what ChatGPT tells your buyers, you have to ask it, and more than once.

1. Write down 20 to 30 questions a buyer asks before they know your name: the best tool for a job, alternatives to a competitor, how to solve the problem you solve. Leave out questions that contain your brand. An engine names you when the question does, so such questions measure nothing. Our two brand questions named us in 48 of 48 answers. 2. Pick the engines your buyers use. For most markets that means ChatGPT, Google AI Overviews and AI Mode, Gemini and Perplexity, then Claude and Microsoft Copilot. Three is a reasonable start. 3. Ask every question more than once. The same question to the same engine gives a different answer on a different day, sometimes with a different winner. The chart at the top of this article is one question, asked eight times. 4. Record four things per answer: named or not, linked or not, position in the answer, and every other brand it names. Keep the full answer text. 5. Turn the answers into a number the right way. Visibility is the answers that name you divided by all answers. Add up first, divide once. Do not average daily or per-engine percentages, because a day or an engine with fewer answers would then count as much as the rest. 6. Read a sample of the answers behind the number. An empty or cut-off answer still counts as an answer, one that named nobody, and it pulls every brand down. We found exactly this in our own data while writing this article. 7. Repeat every week and compare months, not days. A spreadsheet with one row per answer is enough to start. These three rows are real, from 30 September 2026.

We sell an AI tracking tool, so we track ourselves. Our project has 24 questions. Two contain our name. The other 22 are what a buyer of a tool like ours would ask, from "best AI brand visibility tracking tool" to "how do I track my brand's visibility in Google AI Mode and AI Overviews". Three engines answered each question on eight days between 5 and 30 September 2026. Asked the 22 questions, ChatGPT named us in 0 of 176 answers and did not link to our site once. All 13 links to crunchjunkie.io that month came from the two questions with our name in them. Claude and Gemini did not name us in any of their answers either. ChatGPT did name other tools, and often.

The top four are close. With 176 answers the margin of error on each share is about seven points either way, so the order of Profound, Otterly AI, Ahrefs Brand Radar and Scrunch AI is not settled. Our zero is settled. And a tool ChatGPT sometimes put first is not in the table at all, which the next section explains. The table uses ChatGPT only. In the same month 20 of Claude's 176 answers to these questions came back empty, and many of Gemini's stopped after a few sentences: 62 of 176 were shorter than 300 characters. An empty or cut-off answer counts as one that names nobody, so including them would push every brand down. Every ChatGPT answer had text; the shortest was 1,171 characters long. What we take from it: being found when someone asks about you is easy. Being named when someone asks about the category is a different problem, and a harder one. We have work to do.

Record every brand an answer names, not only the ones on your list. Your list will be wrong. Ours was. Semrush was not on our competitor list, so the table above has no row for it. Yet ChatGPT's answer to "best AI brand visibility tracking tool" on 30 September opened with "Best overall: Semrush AI Visibility Toolkit", as it had on 28 September. And semrush.com was the site ChatGPT linked most often: in 84 of the 176 answers. Then look at what the answers link to. For our 22 questions, the ten sites ChatGPT linked most often were all first-party: a vendor's own pages or a platform's own documentation. When an engine needs a fact about a product, it reads the product's own site. A plain page with your prices, features and the engines you cover is the cheapest fix on this list. We looked at this from the other side in the pages AI already cites. Share of voice, your mentions as a share of all brands' mentions, is only as good as the list behind it. Leave a brand out and everyone else's share goes up.

Both are free, and both are worth setting up, as long as you know what they cannot see. GA4 has a default channel called AI Assistant. Google describes it as "the channel by which users arrive at your site from sources like ChatGPT, Gemini, Deepseek, Copilot, or Grok" and adds: "It excludes Google's AI Overviews and AI Mode." There is nothing to install. Open Traffic acquisition and switch the dimension to Session source to see which assistant sent each visit. Search Console has a Generative AI performance report, which Google rolled out to all sites on 31 August 2026. It counts impressions, in Google's words "how many times links to your site were shown to a user in a generative AI feature", in AI Overviews and AI Mode. An empty report can mean you have too few impressions to show, not that you have none. We went through it in Search Console's Generative AI report. Neither tool sees an answer that names you without a click, or one that names a competitor instead. That is the part AI tracking covers.

The method is the same. The details differ by engine. Google's AI Overviews appear only on some searches, so a check is a Google search that may or may not show one. AI Mode is a conversation and behaves more like ChatGPT. ChatGPT and Perplexity answer in an app and through an API, and the two do not always agree; we compared them in scraping vs API. The engines also disagree with each other about the same brand, so one engine is not a sample of AI search. Google's two surfaces have their own guide, tracking AI Overviews and AI Mode, and why AI engines disagree covers the rest.

More than most dashboards show. ChatGPT's first pick for one question changed between four tools in eight answers. Across all 22 questions Profound was named in 74 of 176 answers, 42 per cent, and Otterly AI in 69, 39 per cent. The margin of error on each is about seven points, so that gap could be noise. Two habits keep the numbers honest. Put the number of answers next to every percentage. And treat a change as real only when it is bigger than the margin of error, which on a few hundred answers is several points. We ran the same study twice to see how much rankings move when nothing has changed: noise or signal. Which metrics are worth reporting is in AI visibility metrics that matter.

Doing this by hand works for a month. After that the asking, counting and reading add up to hours every week, and that is when a tool earns its money. Check five things before you buy: which engines it asks and how, in the app or through an API; how often it asks each question; whether it shows the number of answers next to every percentage; whether it records every brand an answer names or only the ones you entered; and whether you can open the answer behind every number. We build one of these tools, CrunchJunkie, so read this section with that in mind. Our buyer's guide to AI search monitoring tools compares the others, with prices. Claude made the same point in one of the answers we tracked: "Most 'best of' lists online are written by vendors who rank themselves first, so I'd treat those rankings with caution." That includes ours.