How to track brand mentions in AI search

The two kinds of mention, and why they need different fixes

Before you count anything, you need to know that there are two different things people call “a mention,” and they are not the same event.

The first is being named from memory. The assistant produces your business name out of what it absorbed while it was being built, or out of what it has read about you elsewhere. Your website was not necessarily opened. It may not even still exist in the form the model remembers.

The second is your page being opened and used as a source. The assistant went out, fetched a page from your website, and built its answer partly from what was on it — then showed a link or a numbered footnote pointing back at you.

It is the difference between a neighbour recommending your bakery from memory and a neighbour reading your menu aloud off your website. Both are worth having. Only the second one had to open your website.

Step 1: write down the ten questions your customers actually ask

Not keywords. Sentences, in the words a real person types into a chat box. Pull them from three places:

• Category questions — what someone asks when they know what they need but not who from. “Best sourdough bakery in Leeds.” “Who should I use for commercial pest control in Manchester?”
• Problem questions — what they ask before they know what to buy. “My basement smells damp after it rains, what do I do?” These matter more than owners expect, because the assistant picks the category on the customer’s behalf and then names someone in it.
• Name questions — “Is [your business] any good?”, “[your business] vs [competitor]”, “[your business] reviews.” These test what the assistant says about you once someone has already heard of you.

Your sales inbox and the questions you answer on every phone call are the source. Write them the way a customer would, clumsy phrasing included.

Ten questions is the working number: enough to show a pattern, few enough that you will actually re-run them. Five is a fine start.

Then freeze the list. Changing the wording changes the answer, so an edited list makes this month and last month incomparable. Date the list at the top. When you think of a new question, add it in a new block below — never rewrite an old one.

Step 2: ask each question five times, in a fresh chat

Pick the assistants your customers actually use — ChatGPT, Perplexity, Gemini, Copilot. Treat each one as a separate result. Never average across them: they are different systems reading different things, and a good score in one tells you nothing about another.

Three rules make the answers worth recording:

1. Start a new chat every single time. A follow-up question inherits everything said earlier in that conversation, including your business name. That contaminates the result.
2. Get out of your own account where you can. These assistants have memory — they carry things about you from one conversation to the next — and yours knows you own the business. Use a temporary or incognito chat, or a logged-out session where the assistant offers one, so you see something closer to what a stranger sees.
3. Never ask a leading question. “Do you know [your business]?” will get you a yes about almost anything. Ask the question a buyer would ask, and see whether you turn up on your own.

Five runs per question, per assistant. Five questions across two assistants is fifty prompts, which is about an hour of copying and pasting. That is the real cost of the manual method, stated honestly.

Why one check tells you nothing

This is the part most people skip, and it is why most of what owners believe about their AI visibility is wrong.

These systems do not look up a stored answer. They generate one, sampling among plausible continuations as they write, so the same question asked twice can produce two different answers with nothing whatsoever changing on your website. When the assistant also goes out and fetches pages, that adds a second layer of variation on top: what it fetches depends on phrasing, timing, and sometimes where you are.

So a single check is closer to a coin flip than a measurement. Named once is not “we are visible.” Missing once is not “we are invisible.”

Named in 3 of 5. Linked in 1 of 5. Competitor B named in 5 of 5. Any single one of those five runs, taken alone, would have told you a different story — and run 3 on its own would have told you that you were winning.

Write the result down as a fraction with the bottom number showing: 3 of 5, not 60%. A percentage computed over five runs is a small fraction wearing a costume.

Step 3: record it in a spreadsheet

One row per run. Not per question — per run. Collapsing five runs into one row throws away the whole month’s work, because the variation is the finding.

One formula does the counting. If the five runs for a question sit in rows 2 to 6 and Named is column E, then =COUNTIF(E2:E6,"yes") gives you the top half of your fraction. Copy it down for each block of five.

Keep a second tab as the monthly summary: one row per question per assistant, with named X of 5, linked X of 5, and the business that appeared most often against you. Keep every raw row. The raw rows are what let you tell a real change from noise six months from now.

The “who else” column is the most valuable one on the sheet. It is not a report, it is a to-do list.

Step 4: read the sheet and pick the fix

Four patterns come out of a filled-in sheet, and each one points somewhere different:

• Named often, almost never linked. The assistant knows your name but is not opening your site. Look at access first: whether automated visitors are allowed in by your robots.txt (a plain text file at the root of your domain that tells automated programs which pages they may read), whether your pages show their words without needing JavaScript to run, and whether a page answering that exact question exists on your site at all. More blog posts do not fix a locked door.
• Linked sometimes, rarely named. You get fetched when a question is specific, but you are not part of the assistant’s default picture of your category. The fix is presence in the sources these systems already read — and your sheet just listed them, in the links that appeared alongside you.
• Neither, while a competitor is both. Read the answer itself. The reason it gives for recommending them is your brief. Most often it is that they have a page answering that question and you do not.
• Named and linked. Good, and quietly perishable. Keep re-running it.

Then work like an experiment: change one thing, wait, and re-run the same frozen question list. Two changes at once and you have learned nothing about either.

Hold this next part firmly, because it is what separates a method from a sales pitch. No engine publishes what makes it name one business rather than another. Every fix above is a reasonable hypothesis that your sheet then tests. Nothing on this page, or on anybody else’s, makes a mention certain.

What you cannot track, and why

An honest tracking routine has holes in it. Here are yours, named rather than papered over.

• Google’s AI Overviews and AI Mode. There is no public API for them — an API being a direct connection a program can use to ask a system a question and get a machine-readable answer back. What appears is also shaped by who is asking and where they are, so two people see different things. Look at them by hand if you like, but label that a spot-check, and never turn spot-checks into a rate.
• Assistants inside apps and shopping surfaces. Amazon’s shopping assistant, and the assistants built into messaging apps, have no route for repeatable checking at all. Anything you read reporting on them is a person looking, or a scrape.
• Anything behind a login. A company’s internal assistant, or a tool inside someone’s workspace, cannot be observed from outside it.
• How many people asked. Nobody publishes how often a given question is put to an assistant. Any “prompt volume” figure you are shown is modelled, not counted. So you cannot rank your ten questions by demand — you rank them by your own judgement about your buyers, which is a weaker instrument, honestly labelled.
• Why you were or were not named. You can see the outcome. You cannot see the rule.
• Last month, if you did not measure it. There is no archive of past answers to go back to. Your measurement starts the day you start.

Where a tool helps, and where it doesn’t

A tool buys you three things and no more: the repetition, so fifty prompts happen without your afternoon; the history, so a trend exists instead of a memory; and a neutral account, so answers are not shaped by your own chat history. It also checks for the link rather than relying on your eyes at prompt forty-eight. What it does not buy you is knowledge of the ranking factors, because nobody has those.

Our own AI visibility checker is in build and not open yet, and the free tools shelf goes up one tool at a time as each ships. Until then the spreadsheet above is the same measurement done by hand. Slower. Not less honest.

The sheet is not the deliverable. The name in the “who else” column is.

FAQs