How to track brand mentions in ChatGPT, Perplexity and Gemini

By Miles Burke, founder of CitableReport. Published 15 October 2026.

To track brand mentions in ChatGPT, Perplexity and Gemini, you put the same set of customer questions to each assistant on a regular schedule, and record three things every time: whether your brand is named, whether your site is linked, and which competitors appear instead.

Because AI answers change from one run to the next, the number that matters is how often you appear across many answers, not where you appear in any single one. This guide sets out the method step by step, with a tracking template to download.

Why is tracking AI mentions different from tracking search rankings?

Tracking AI mentions is different because there is no fixed position to track. A search results page is much the same for everyone who searches the same words on the same day. An AI answer is written fresh each time, and the brands in it change.

SparkToro measured how much. In research published in January 2026, volunteers ran 12 prompts through ChatGPT, Claude and Google's AI a combined 2,961 times, and found that ChatGPT and Google's AI had under a 1 in 100 chance of giving the same list of brands twice; Claude was only slightly steadier. The order changed even more.

What did hold steady was which brands turned up at all: some appeared in most answers to a question, however the list was shuffled. The researchers' conclusion was to measure visibility, the share of answers that mention you, and to distrust any tool that reports a "ranking" in AI.

People also ask the same thing in very different words. A companion SparkToro study found that prompts written by different people for the same need had almost nothing in common, yet the assistants still tended to name a similar set of brands.

So a tracking set should cover the needs your customers have, not one perfect phrasing of each.

Step 1: Choose the questions to track

Start with 20 to 50 questions your customers really ask, split across three kinds:

• Category questions, where you want to be recommended: "What is the best invoicing app for a sole trader?" or "Which accountants in Brisbane specialise in medical practices?"
• Problem questions, where your expertise should be quoted: "How do I chase an unpaid invoice politely?"
• Brand questions, where the facts about you should be right: "What is [your brand]?", "Is [your brand] any good?" and "[your brand] or [competitor]?"

The best sources for them are the words customers already use: the search terms in Google Search Console, questions in sales emails and support tickets, and the phrases in your reviews.

Write them the way a customer would type them, including the casual ones. Once you have the set, keep it fixed for at least six months, or the months will not compare.

Step 2: Decide which assistants to check, and how

Check the assistants your customers are likely to use: usually ChatGPT, Gemini, Perplexity and Google's AI Mode, plus Claude and Microsoft Copilot if your audience is technical or works in offices that use Microsoft 365.

Then make each test as neutral as you can, so you see what a stranger sees rather than what the assistant has learned about you:

• In ChatGPT, use a Temporary Chat, which by default does not use your saved memories or custom instructions, and make sure web search is on.
• For the others, use a private browser window, signed out where the assistant allows it.
• Ask from the country your customers are in. Answers about local services change with location.
• Note the date, the assistant and the mode you used, because the models behind them are updated without notice.

Step 3: Ask each question more than once

Ask each question several times, because one answer is a sample, not a result. SparkToro suggests that a reliable picture of visibility for a single prompt takes at least 60 to 100 runs, which is more than anyone will do by hand. A practical compromise is three runs of each question on each assistant every month.

That will not give you a precise percentage for any one question, but across 20 or more questions the overall share is steady enough to show a trend. If you need precision for a handful of questions that matter most, that is where an automated tool earns its fee.

Step 4: Record what each answer says

For every answer, write down:

• Whether you are mentioned, by name, anywhere in the answer.
• Whether you are linked: is your site one of the sources, or a citation, or only a name in the text?
• Which competitors are named, so you can compare.
• Which sources are cited. Perplexity links every answer to its sources, and ChatGPT and Gemini usually do when they search. These show you which sites the assistant trusts on your topic.
• Whether what it says about you is right: prices, locations, what you do. Wrong facts are worth fixing at their source.

A spreadsheet is enough. Our AI brand mention tracker template (CSV, opens in Excel, Numbers or Google Sheets) has a column for each of these, with two example rows to show how it is filled in.

Step 5: Work out your share of mentions

Your share of mentions is the number of answers that name you, divided by the number of answers you collected. Work out the same figure for each competitor and you have a comparison that means something.

For example, 20 questions asked three times on each of three assistants gives 180 answers. If you are named in 27 of them, your share of mentions is 15%. If your main competitor is named in 63, theirs is 35%.

Do the same for links: named and linked in 9 answers is a citation rate of 5%. Split the figures by question type too, and look hardest at category questions: appearing when someone asks for you by name is good, but category questions are where new customers come from.

Step 6: Check the numbers against real visits

Mentions are a means to an end, so check whether they turn into visits. In your analytics, set up a view of visits referred by chatgpt.com, perplexity.ai, gemini.google.com, claude.ai and copilot.microsoft.com. OpenAI's publisher FAQ says ChatGPT adds utm_source=chatgpt.com to the links it gives, which makes its visits easy to separate.

For Google, clicks from AI Overviews and AI Mode are counted within ordinary web search figures in Search Console, so watch the pages that answer your tracked questions. If your share of mentions rises and those visits do not, look at whether you are being linked or only named.

Should you use a tool to track AI mentions?

A tool is worth it once you track more than a few dozen questions, or need the 60 to 100 runs per question that a precise figure takes. Several SEO suites now offer AI tracking, and there are dedicated tools too. Whichever you look at, check that it:

• reports how often you appear across many runs, not a single "position" in an AI answer;
• shows the sources each answer cited, not just whether you were named;
• lets you read the answers themselves, so you can check its counting;
• says how it asks the questions. Whether answers fetched through a developer API match what people see in the apps is still an open question, which the SparkToro researchers raised for further work.

What should you do if AI assistants do not mention you?

Start with the sources column. If the same review sites, directories, forums or publications keep being cited for your category, being present and well described on those sites is the most direct way into the answers.

Then check your own site: if AI crawlers cannot get in, or your pages do not answer questions clearly, you will not be cited even when you are the best answer. Our guide to doing an AI visibility audit covers that step by step, and what an AI visibility audit is explains why each part matters.

One thing to be clear about: CitableReport does not track mentions. It checks your website, the part of AI visibility you control directly, and its report includes the measurement questions above as a checklist.

Tracking tells you whether you are mentioned; an audit tells you what on your site may be stopping it. Most businesses need both.

How often should you track brand mentions in AI?

Track monthly. Weekly is too often for a hand-run process, because the run-to-run noise will be bigger than any real change, and quarterly is too slow to see whether a fix worked.

Run an extra round after anything likely to move the answers: a product launch, a burst of press coverage, a new set of reviews, or a change to your site. Keep every month's sheet. The trend over six months is the thing worth reporting.