How to track brand mentions in AI search, by hand or with a tool

How to track brand mentions in AI search: the short answer

You track brand mentions in AI search by asking the same questions again and again, and counting how often the answers name you. There is no ranking to look up and no report from the assistants themselves, so the only record is the one you build.

Think of it like a mystery shopper. You cannot see what every buyer was told, so you send the same shopper in with the same questions every week and write down what they hear. The questions stay fixed so that a change in the answers means something.

The method has four parts. You choose the questions. You ask every assistant every question. You keep the full text of every answer. Then you count, with the denominator written next to every figure. The rest of this guide takes each part in turn, then covers what tools add and where they can mislead you.

A brand mention figure is only as honest as its denominator. Write it as named in 6 of 120 answers, never as a bare percentage.

What counts as a brand mention in an AI answer

A mention is an answer that writes your brand name in its text. That is different from a citation, which is a link to one of your pages, and different from a recommendation, where the answer tells the buyer to choose you.

Keeping these apart matters because they move separately. An answer can link to your blog post as a source and never write your name. Another can name you in a list of ten and steer the buyer to a rival in the next sentence. Tools differ on which of these they count, so check before you compare two figures.

Most people start with mentions, because they are the easiest to count and the closest to the question that worries them: does the assistant bring us up at all? Citations are covered on the citation tracking page, and claims in our guide to AI brand sentiment.

Why one check is not enough

One check is not enough because assistants write each answer fresh, and the brands they name change from one try to the next. A single answer is one roll of the dice, not a measurement.

The clearest public evidence is a study by SparkToro and Gumshoe, published on 28 January 2026. Volunteers ran 2,961 prompts across ChatGPT, Claude and Google's AI results. The study found a less than 1 in 100 chance that ChatGPT or Google's AI gave the same list of brands in any two responses, and less than 1 in 1,000 that the order matched (SparkToro).

The same study found something steadier underneath. How often a brand appeared across many runs was fairly stable. One hospital appeared in 69 of 71 ChatGPT responses to its question, and the top headphone brands appeared in 55% to 77% of responses (SparkToro). The authors concluded that visibility across dozens to hundreds of prompts, run many times, is a reasonable metric.

That shapes the whole method. Do not track where you appear in the list, because the order is close to random. Track how often you appear, over enough answers that one lucky or unlucky answer cannot swing the figure.

How to track brand mentions by hand, step by step

You can do this with a spreadsheet and an afternoon a week. It is slow, but it teaches you exactly what a tool should be doing, which makes you a much harder buyer to fool later.

1. Write 20 to 50 questions a buyer would really type, in their words, not yours. Our guide to finding buyer questions shows where to look.
2. Mark each question as branded if it contains your name, and non-branded if it does not. Keep at most a small share branded.
3. Pick the assistants your buyers use. Today that usually means ChatGPT, Perplexity and Gemini, plus Google's AI results if search matters to you.
4. Use a fresh session with no memory or custom instructions, so your own history does not shape the answer.
5. Ask every question on every assistant, and paste each full answer into your sheet with the date and the assistant's name.
6. For each answer, record whether it named you, which other brands it named, and which sites it linked to.
7. Count the non-branded answers that named you, over all non-branded answers, and write both numbers down.
8. Repeat with the same questions on the same day each week, and change nothing about the questions between weeks.

The sheet itself can be simple. The table below shows one row per answer, which is the shape that lets you recount anything later.

Here is how the count reads with made-up numbers. Say you ask 30 non-branded questions on four assistants, which gives 120 answers. If 9 of them name you, you write named in 9 of 120 non-branded answers. You do not write 7.5% on its own, because the reader cannot tell whether that came from 12 answers or 1,200.

Which assistants to check, and how each shows its sources

Check the assistants your buyers actually use, and record each one separately, because they search different sources and name different brands. A pooled figure across all of them hides an assistant where you never appear.

ChatGPT, Perplexity, Gemini and DeepSeek each answer from a mix of what the model learned in training and what it finds by searching the web at the time. The balance differs by assistant and by question. Our engine pages describe what is publicly documented about each one.

Google's own AI results are a special case. Google says that sites appearing in AI features such as AI Overviews and AI Mode are included in the overall search traffic in Search Console (Google Search Central). That means Search Console will not separate them for you. You still have to look at the AI result itself to know whether it named you.

Whatever you check, keep the list the same from week to week. Adding an assistant halfway through changes the denominator, so a rise or fall that week is partly the new assistant and not a change in what buyers see.

What tools do, and what to ask a vendor

A tool does the asking and the counting for you, on a schedule. The useful question is not which tool has the biggest number, but whether you can see the questions, the answers and the denominator behind every figure it shows.

Tools differ a lot in where their questions come from. Ahrefs says Brand Radar takes People Also Ask questions from its keyword index and enters them into the web version of each chatbot using the default model (Ahrefs help). That gives a large shared sample, but the questions are not ones you chose for your buyers.

Tools also differ in what they call a mention. Peec AI's documentation defines visibility as the percentage of AI responses where your brand appears (Peec AI docs). Conductor's documentation describes brand mentions as how often your brand name, and variations you define, appears directly within the text of AI responses (Conductor docs). Both are per answer counts, but the variations you define can change the result.

Before you pay for any tracker, ask these questions and expect a plain answer to each:

• Can I see and edit the exact questions that are asked?
• Can I open the full answer behind any figure, word for word?
• Does every figure show both counts, not only a percentage?
• Are questions that contain my name reported apart from the rest?
• Does it show how much the figure moves when nothing has changed?
• Which assistants are asked, and through which route: the public web app or an API?

The last question matters because an assistant reached through its API can behave differently from the app a buyer uses. Neither is wrong, but a vendor should say which one it measured.

Where Agentcites fits

Agentcites is one tool that does this tracking, and it is built around the method above. It reads your site, writes up to 50 buyer questions for you to review, and asks them of ChatGPT, Perplexity, Gemini and DeepSeek every week. It keeps every answer word for word and reports questions with your name apart from the ones without it, as the answer tracking page describes.

It has limits worth knowing before you choose it. It only counts brands on its tracked list. It adds a rival to that list when the rival keeps winning your questions, but a brand that wins once is visible only through the question you lost. It asks each question once per scan, so it does not yet tell you on its own whether a small change is real. It costs $149 a month per website, and the pricing page lists what that includes.

Mistakes that make the count lie

Most bad brand mention figures come from a handful of mistakes, and every one of them makes the figure look better than it is. Check for these before you report anything.

• Mixing in branded questions. An assistant asked about your brand by name almost always names you back, which inflates the figure without a single new buyer finding you.
• Reporting a rank. The order of brands in an answer changes almost every time, so an average position is mostly noise.
• Changing the questions. A new question set is a new measurement, not a trend, so start a fresh baseline.
• Counting from too few answers. With 12 answers, one answer moves the figure by more than 8 points.
• Throwing the answers away. Without the text, nobody can check a figure or recount it a better way later.
• Pooling every assistant. A healthy total can hide one assistant that never names you at all.

Our guide on reporting AI visibility honestly explains these rules in more detail, with worked examples of each.