How to track brand mentions in ChatGPT (without fooling yourself)

Tracking brand mentions in ChatGPT means asking the assistant the questions your buyers ask, without naming your brand in them, then recording how often the answer names you anyway. Each answer is one sample, not a ranking position, so the output is a rate with a margin of error rather than a number.

Widely read guides on this topic are largely about which dashboard to buy and how often to send the client a report, and the ones we read said nothing about how many times to ask, what to do with an answer that refuses, or why the same question returns a different brand list the next day. That gap is what decides whether your number means anything, so it is what this page covers.

You can run the whole method in a spreadsheet. It costs you an afternoon, not a subscription.

1. Write the questions your buyers ask, not the ones about you

The most common way to break this on the first day is to ask What do you think of <brand>?. The assistant will discuss your brand because you handed it the name. You have measured your own question.

Write the questions the way somebody who has not heard of you would type them:

• "Which HR software would you recommend for a small company in Germany?"
• "Best project management tools for a five-person design studio"
• "Who builds one-page landing pages with conversion tracking already set up?"

Three rules keep the sample honest.

Keep your brand name out of the question. If the name is in the prompt, the mention belongs to you, not to the engine. Our own question generator enforces this in code: any question other than a deliberate vendor_check is rejected if it contains the brand or one of its aliases, and vendor_check answers are kept out of the headline rate. The full gate is written out on the measurement method page.

Keep the buying context in. Location, company size, budget, the constraint that actually decides the purchase. A question with no context returns a generic list that no real buyer would have received.

Work out the number of questions from the interval you need, not from a rule of thumb. Questions multiplied by engines multiplied by repeats is the denominator behind your mention rate, and that denominator is what decides how wide the interval comes out. Ten questions asked three times on one engine is thirty answers, and a perfect thirty out of thirty still only bounds the true rate above 88.6%. Twenty questions is sixty answers, and a perfect sixty out of sixty bounds it above 94.0%.

Pick the width you can act on and read the size off that. A longer list is not better than more draws on the right questions, and it costs more to maintain when you have to keep it identical week after week.

2. Ask each question more than once

The same question asked twice does not return the same answer. That is how these models behave, not a fault in your setup.

Ask each question at least three times, each in a fresh conversation. If you are doing it by hand in the ChatGPT app, open a new chat every time, because earlier turns in the same thread steer the next answer. Three is the cheapest number that shows you whether the answer moves at all. It is not enough to call a small week-over-week change real, and nothing in this method makes it enough.

We use three repeats per question per engine on paid monitoring for exactly that reason, and the method page says plainly that this is a cost decision rather than a statistical one.

3. Record four things per answer

One row per answer, not per question. If you asked ten questions three times, you have thirty rows.

Field	What to write down	Why it earns its column
Mentioned	Whether the answer named your brand in its prose. Yes or no	This is the numerator of your rate
Position	If the answer is a ranked or bulleted list, the place your brand sits in it	Being named ninth is not the same result as being named first
Cited sources	Every URL the answer points to, and whether the domain is yours	A citation and a mention are separate events that can happen without each other
Competitors	Every other brand named in the same answer	Tells you who holds the answer you want, which is where the diagnosis starts

Two details decide whether these rows are comparable later. Strip link targets and bare URLs out of the answer text before you look for your name, or a domain sitting inside a link will be counted as your brand being named in prose. And record refusals as refusals rather than as a no, because an engine declining to recommend anyone is not a fact about you.

The competitor column is the one people leave out on day one and wish they had by week four.

4. Turn the tally into a rate you can compare

Your headline number is the mention rate: valid answers that named you, divided by valid answers. Valid means the answer neither errored nor refused; refusals are counted and shown on their own line rather than folded into the rate. Those definitions match the ones on the method page, and if you want to compare your figures to ours, they need to.

A bare percentage from thirty samples invites you to read noise as a trend, so put a range on it.

The Wilson score interval is the standard choice for a proportion from a small number of trials. It was developed by E. B. Wilson in 1927, and unlike the textbook normal-approximation interval it "doesn't suffer from problems of overshoot and zero-width intervals" (Wikipedia, Binomial proportion confidence interval, retrieved 2026-09-20). Those two failures are exactly the ones you hit here: at 0% or 100% the normal interval collapses to nothing and tells you that you are certain, which you are not.

A worked example from one of our runs

On 2026-09-20 we ran our free-check path against the ChatGPT API: three buyer questions for the US market in English, generated for the brand Calendly in meeting scheduling software, each question asked once. That is three draws in total. It is a smoke test, not a baseline, and the arithmetic below is the reason we describe it that way. The raw output is recorded in full.

The three questions, verbatim:

1. "Can you recommend meeting scheduling software that works well for remote teams in the US and supports multiple time zones?"
2. "Which meeting scheduling tools are best for small businesses in the US with a limited budget under $20 per month?"
3. "How do the top meeting scheduling software options compare in terms of ease of use and integration with popular video conferencing platforms?"

Calendly was named in all three answers, so the tally is 3 out of 3 and the point estimate is 100%. The other brands the extractor recorded across those same three answers were Doodle (twice), and SavvyCal, Reclaim.ai, Morgen and Square Appointments (once each).

Now the interval, with p = 1, n = 3 and z = 1.96:

centre = (1 + 1.96² / 6) / (1 + 1.96² / 3)
       = (1 + 0.6403) / (1 + 1.2805)
       = 1.6403 / 2.2805
       = 0.7192

half   = 1.96 · √( 0 + 1.96² / 36 ) / 2.2805
       = 1.96 · 0.3267 / 2.2805
       = 0.2808

Wilson 95% interval: 43.8% to 100%.

Read that twice. A brand that ChatGPT named in every single answer we drew is still statistically compatible with a true mention rate below one in two. Three draws cannot distinguish "always mentioned" from "mentioned slightly more often than not". If we had instead seen zero out of three, the interval would have run from 0% to 56.2%, which is compatible with a brand that gets named most of the time. Both of those rows are in the table on the method page.

The practical rule: if this period's interval overlaps the last one, you have not measured a change. Keep the question set identical between runs, or you have two snapshots of two different things and no trend at all.

The four mistakes, and how you notice you are making one

1. The brand name is in the question. How you notice: your mention rate sits at or near 100% from the first run and never moves, in a category where you know you are not the obvious default. Read your own prompts back. If your name is in them, the rate is measuring your typing.

2. One draw per question. How you notice: you re-ran a question out of curiosity and got a different set of brands, or your weekly line zigzags while nothing on your side changed. A sample size of one has no variance to report, which is not the same as having no variance.

3. Citations counted as mentions. How you notice: two people on the team tally the same answer and disagree, or your number jumps when you switch between reading the answer body and reading the source list. Give them separate columns and never add them together.

4. The question set drifted. How you notice: your sheet has no column holding the exact question text, so you cannot reconstruct what produced last month's figure. If you have to reword a question, treat it as a new question with its own start date rather than editing the old row.

What this does not cover

Google AI Overviews and AI Mode are outside what we measure. Our monitored surfaces are ChatGPT, Gemini and Perplexity. A number produced by this method says nothing about what Google shows above the blue links, and should never be presented as if it did.

Personalisation. These measurements are taken through APIs with no signed-in user, no chat history and no app context. The result describes what the engine says to an anonymous caller. An individual customer with two years of chat history may see something else, and this method has no way to observe that.

Cause. The mention rate is a correlation with whatever else was true that week. It tells you where you stand, not why you stand there, and a rate that moves after you publish something is not evidence that the publishing caused it. The competitor and cited-source columns are where a hypothesis about cause starts; confirming one needs a deliberate before-and-after with intervals that do not overlap.

Small samples. Nothing here removes the width of the interval. A short question set measured occasionally cannot detect a small change, and the honest report says inconclusive.

Doing it on a schedule

Manual measurement works until the week you are busy and skip it, and a trend line with a hole in it is not a trend line. If you want it running without you across three engines with the same questions every week and the intervals calculated for you, that is what we sell: see what AI visibility tracking measures, or run the free check on one brand to see the shape of the output first. Bear in mind that the free check is three questions asked once each, which is the same three draws as the example above, with the same wide interval.

If you would rather keep it in a spreadsheet, everything above is the whole method. None of it needs our tooling.

Questions

How do I check whether ChatGPT mentions my brand? Write questions a buyer in your category would ask, none of them containing your name, ask each one at least three times in a fresh conversation, and record whether your brand appears in the answer prose. The share of valid answers that named you is your mention rate. How many questions you need depends on how narrow an interval you want; the arithmetic is in the section above.

How do I monitor brand mentions in ChatGPT over time? Keep the question set frozen and repeat it on a fixed schedule, weekly or monthly. Compare the confidence intervals rather than the point estimates. Changing the wording of a question resets its history, so give any reworded question a new row and a new start date.

How do I see which competitors ChatGPT recommends instead of me? Record every other brand named in each answer, in the same row as the mention. After a few runs the tally shows which names hold the answers you want. In our 2026-09-20 run on meeting scheduling software, the extractor recorded Doodle twice and SavvyCal, Reclaim.ai, Morgen and Square Appointments once each across three answers.

How do you measure LLM visibility? As a rate with an interval, not as a score. Mention rate is the headline; average position, citation rate and share of voice answer different questions and are defined on the method page. Any single composite "visibility score" that hides its denominator is not something you can argue with.

How do I monitor LLM SEO performance? Measure the same question set before and after the change you made, and check whether the two intervals overlap. If they do, the work has not yet shown an effect at this sample size, which is different from the work having failed. This is also why the question set must not change at the same time as the content does.

How many questions and how many repeats do I need? We do not publish a recommended number, because we have no data that would support one. What we can give you is the arithmetic: questions multiplied by engines multiplied by repeats is your denominator, and the denominator sets the width of the interval. Thirty answers with a perfect result bounds the true rate above 88.6%; sixty answers bounds it above 94.0%; three answers bound it above 43.8%, which is why a free check is a smoke test. More repeats narrow the interval faster than more questions do, because more questions measure more things rather than measuring one thing better. Writing the question set goes through this in full.

Does this work for Google AI Overviews? No. AI Overviews and Google's AI Mode are outside what we measure, and the method on this page is written for engines with an official API that returns the sources behind an answer.

What counts as a mention? Your brand name, or a confirmed spelling of it, appearing in the answer's prose after link targets and bare URLs have been stripped out. A domain that appears only inside a link is a citation, and belongs in the citation column instead. Related reading on what the monitoring covers: AI visibility tracking.