How to Track Brand Mentions in ChatGPT

You asked ChatGPT about your business last week and it named you. You asked the same question this week and it didn't. Is that news?

Usually, no. Sometimes, yes. And the whole skill of tracking brand mentions in ChatGPT comes down to telling those two apart. Running one honest check is the easy part, and we cover it step by step in our guide to checking what ChatGPT says about your business. This post is about what comes after: asking again next week, and the week after, and knowing what the difference between two answers actually means.

That turns out to be a different job from checking. A single check asks "what does ChatGPT say?" Tracking asks "has something changed, and if so, what?" To answer the second question you need three things a one-off check never needs: a way to repeat the check so rounds are comparable, a rule for when a change is big enough to believe, and a log that records enough to tell you why the answer moved.

Why the Answer Changes When Nothing Has Happened

Start with the uncomfortable fact. ChatGPT does not give the same answer to the same question twice, even when nothing about your business, your website or the question has changed.

This isn't a bug that's about to be fixed. Researchers at Thinking Machines, who study exactly this, open their write-up with an everyday example: "you might observe that asking ChatGPT the same question multiple times provides different results" (Thinking Machines). Some of that comes from deliberate randomness built into how these models pick words, and some of it comes from technical details of how the servers run, which the same researchers show can change an answer even when the randomness is turned off.

For lists of businesses, the variation is large. In SparkToro's January 2026 study, 600 volunteers put the same questions to ChatGPT, Claude and Google's AI thousands of times, and the authors found that the AIs "rarely give the same list of brands or recommendations twice (<1 in 100 times, no matter the question)" (SparkToro).

So think of each answer as one roll of a weighted die. If ChatGPT names you in about three answers out of five for a given question, then on any single ask you'll either be named or not, and which one you get is partly luck. One answer can't tell you the weighting. Only many answers can.

That reframes what you're tracking. You are not tracking "does ChatGPT name us?", which has no stable yes or no answer. You are tracking how often it names you, and whether that rate moves.

Turn Each Round Into a Rate, Not a Yes or No

The single most useful change you can make to a tracking routine is this: never ask a question once per round. Ask it several times, each in a fresh chat, and write down how many of those answers named you.

A round then produces "3 of 5" instead of "yes." That one change makes everything later in this post possible, because a rate can be compared and a single yes can't.

How many times is enough? SparkToro's own advice for a reliable number is to run a prompt "usually at least 60-100X, then average these out." That is the right figure for research, and far more than anyone can type by hand every week. For a hand-run routine, five asks per question per round is a workable compromise. It is not precise. It is enough to separate "almost always" from "sometimes" from "almost never," which is the distinction you care about most. The next sections show how to read a five-ask rate without fooling yourself.

The rules that keep the asks comparable are the same ones that make a single check honest. Each ask goes in a new temporary chat set to unpersonalized, with Search turned on, so your own account history isn't steering the answer. Our one-off check guide explains each of those settings. Tracking adds just one more rule on top: freeze the exact wording of every question. Write each one down once, give it a short label (Q1, Q2 and so on), and paste it from your notes each time instead of retyping it. Two slightly different wordings are two different questions, and a change between them tells you nothing.

The Four Reasons a Result Changes

When a rate moves between rounds, there are only four places the change can have come from. Every column in a good tracking log exists to tell these four apart.

1. Chance. The die rolled differently. Nothing changed anywhere. This is the most common cause by far, and the one people most often forget to rule out before acting.

2. Your measurement changed. You worded the question slightly differently, forgot to start a fresh chat, were logged in this time, or the answer came back without any sources because no search ran. The answer moved because the test moved.

3. ChatGPT changed. OpenAI updates ChatGPT constantly: the models behind it, the way it searches, which features are switched on. Its public release notes had eight separately dated entries between 25 September and 8 October 2026 alone. Not every entry touches search or answers, but any of them can, and none of them has anything to do with your business.

4. The web changed. Something ChatGPT reads is different. Either your own site changed (a redesign, a broken page, a page newly blocked from crawlers) or someone else's did (a competitor published a better page, a directory added or dropped you, a review site updated its list).

Only the fourth cause is something you can act on directly, and only half of it is about your own site. Causes one to three are things to rule out, not fix. A tracking habit that skips ruling them out will send you chasing problems that don't exist.

What to Log So You Can Tell the Causes Apart

A tracking log doesn't need to be clever. It needs one row per question per round, and columns that each answer one of the four causes above. Here is a log built for that job:

Column	What goes in it	Which cause it helps rule out
Date	The day of the round	Lines up with release notes and your own site changes (3, 4)
Question label	Q1, Q2... pointing at the frozen wording	Measurement drift (2)
Asks	How many times you asked (5)	Measurement drift (2)
Searched	How many of those answers showed sources	Measurement drift (2): an answer with no sources didn't look anything up
Named	How many answers named your business in the text	The rate you're tracking
Your site cited	How many answers listed a page from your site as a source	The rate you're tracking, second half
Sources seen	The other websites that showed up as sources	The web changed (4)
Site changes	Anything you changed on your site since the last round	The web changed, your side (4)
Notes	Anything odd, plus any ChatGPT release note dated since the last round	ChatGPT changed (3)

Two columns deserve a word.

"Named" and "Your site cited" are separate on purpose. For tracking, what matters is when one moves and the other doesn't. If your name rate holds but your site drops out of the sources, ChatGPT still knows you but has found somewhere else to read about you, so look at what replaced your page. If your site keeps getting cited while your name rate falls, your page is still being read but isn't making clear enough who's behind it.

"Sources seen" is the most informative column in the whole log. It's also the one people skip because it's tedious. When your rate falls, this is where you see why: a new website appearing in the sources week after week, or an old one disappearing. Without it, a drop is a mystery. With it, a drop often comes with a name attached.

This log is built for spotting change. If you also want to catch ChatGPT getting your basic facts wrong, such as your hours, your services or your service area, that's a different kind of check with its own column layout, and our post on why to track at all covers it. Many people run both on the same day from the same answers.

How Many Questions, and How Often

Questions: five to ten is plenty for a small business. Mix them. Include one or two that use your business name ("What does [business] do?"), which test whether ChatGPT can find you at all, and several that don't ("best [service] in [city]", "who can help me with [problem]"), which test whether it recommends you to someone who has never heard of you. The second kind is harder and more valuable, and it's the kind most likely to move.

Rounds: every two to four weeks is a sensible rhythm for a hand-run log. Our breakdown of real AI metrics suggests a monthly look for this kind of check, and for good reason: run it every few days and you mostly measure chance, because ChatGPT's answers don't usually shift for real that quickly. Weekly is fine if a tool is doing the asking for you.

Extra rounds: add one shortly after any deliberate change, whether you rewrote a key page, fixed your business details, or got listed somewhere new. Then add another one a round or two later. A change worth making should show up in more than one round.

What it costs: eight questions, five asks each, is forty asks per round. Done carefully, that's well under an hour, most of it spent copying source lists into the log.

Telling a Real Drop From Noise

Here is the part most tracking advice skips. A rate built from five asks will bounce around a lot even when nothing has changed, and it's worth seeing how much.

Suppose ChatGPT's true rate for one of your questions is three answers in five. Ask five times, and simple probability says what you'll see:

• About one round in three, you'll be named four or five times. It will look like you're doing great.
• About one round in three, you'll be named only two times or fewer. It will look like you dropped.
• Getting named zero times out of five happens only about one round in a hundred.

Nothing changed in any of those rounds. The die simply rolled. Those figures are basic probability for five asks with a true rate of three in five, not a study result. But they explain why a drop from "4 of 5" to "2 of 5" on one question in one round is not something to act on.

So how do you spot a real one? Four tests, roughly in order of usefulness.

Test 1: Did it fall across many questions at once? Chance hits questions one at a time, at random. If one of your eight questions dropped, that's expected. If six of the eight dropped in the same round, something real moved: a change on your site, a change in ChatGPT, or a change in your measurement. Chance doesn't line up like that.

Test 2: Did it last? A real change persists. Chance usually bounces back the next round. Before acting on a single-question drop, run that question again a round later. If it stays down for two or three rounds in a row, believe it.

Test 3: Did it fall to nothing from something steady? If a question named you four or five times out of five for several rounds and now names you zero times, that's very unlikely to be chance (see the one-in-a-hundred figure above). Treat it as real straight away.

Test 4: Did the sources change at the same time? If your rate fell and a new website started appearing in the sources in the same round, that's a strong sign the web changed. If your rate fell and the sources look the same as always, lean toward chance or a ChatGPT-side change.

A worked example

Picture a small bookkeeping firm tracking six questions, five asks each, every two weeks. This is an illustration, not real data. Its "Named" column for one question, "best bookkeeper for small businesses in [city]", reads:

Round	Named (of 5)	Sources seen (new ones in bold)
1	3	Their own site, two directories, a local business association
2	4	Same as round 1
3	2	Same as round 1
4	3	Same as round 1
5	1	Their own site, two directories, a new "top 10 bookkeepers" list
6	1	Their own site, one directory, the new list
7	0	The new list, one directory

Round 3's dip from 4 to 2 fails every test. One question, one round, same sources, and it recovers in round 4. That's chance. Rounds 5 to 7 pass three tests. The drop lasts, it reaches zero from a steady three-ish, and a new source arrives at the exact moment it starts. Two of their other questions dropped over the same rounds. That's a real change, and the log has already pointed at its likely cause: a list article that doesn't include them, which ChatGPT has started leaning on.

Without the "Sources seen" column, the firm would know something went wrong around round 5 and nothing more.

What to Do When the Drop Is Real

Once you're satisfied a drop is real, match the action to the cause.

If your measurement changed: fix the routine, and treat the next round as a new starting point rather than comparing it to the old ones. A broken comparison is worse than none, because it looks like evidence. If you rewrote a question on purpose, give it a new label instead of keeping the old one.

If ChatGPT changed: check the release notes for anything dated between your two rounds that touches search, answers or the default model. If you find one and the drop is across the board, there's usually nothing to fix on your side. Keep tracking. Model changes shuffle everyone, and your rate often settles at a new level within a round or two. If it settles low, look at which sources the new behaviour prefers, which brings you to the next case.

If someone else's site changed: open every new source in your log and read it the way ChatGPT would. Is it a list you could reasonably be on? A directory where your listing is missing or wrong? A competitor page that answers the exact question more directly than yours does? Each of those leads to a specific job: asking to be included, fixing a listing, or improving the page that should be answering that question. Our post on why ChatGPT recommends a competitor walks through what those sources usually are and how to compete with them.

If your own site changed: look at what you changed since the last good round. The usual suspects after a redesign or move are pages accidentally hidden from search engines, broken redirects, business details that went missing from the page, or a template that stopped outputting your structured data. If the drop lines up with a site move, our migration checklist has the order to check things in. A readiness check of the affected pages will show whether AI crawlers can still reach and read them.

If you can't tell: keep tracking for another round before doing anything. Two more data points are cheap. A week spent fixing something that wasn't broken is not.

Gains Need the Same Tests

Everything above applies in reverse. A jump from 1 of 5 to 4 of 5 on one question in one round is as likely to be chance as a drop is. If you made a change and want to know whether it worked, look for the same signs: a rise across several questions, lasting more than one round, ideally with your own pages appearing in the sources more often than before.

This matters most when you're reporting results to someone else. A single good round shown in a meeting as proof that some work paid off is a claim that will probably reverse the next month. Report a gain once it has held for two or three rounds, and say how many asks the rate is built on.

Tracking It Automatically

Doing all of this by hand is realistic for a few questions. It gets tedious fast, and the tedious part, copying sources into a log every few weeks, is exactly what gets skipped first.

We built a version of this into the AI Schema Gen dashboard as AI Visibility tracking, available on Pro plans and above. Here's exactly what it does, so you can judge how it fits the method above.

What it asks: you add questions in a customer's words (5 on Pro, 15 on Premium, 40 on Agency), and a check runs each one about once a week. You can also run a check by hand at any time. The ChatGPT check goes through OpenAI's own search-enabled model with search switched on every time, so every answer comes back with sources and no ask is wasted on a reply from memory. That is close to what the ChatGPT app does when it searches, but it is not literally the app, so your own hand checks and the dashboard won't always match word for word.

What it records: whether a page from your site appeared in the answer's sources, where it came in that source list, every website that was cited, and which of your tracked competitors' sites were cited. It keeps an excerpt of the answer itself. That is the "Your site cited" and "Sources seen" columns of the log above, filled in automatically and kept as a history. It does not separately detect your business name appearing in the text of the answer, so the "Named" column is still worth a hand check now and then.

What it alerts on: an email or in-app alert when your site stops being cited for a question, starts being cited, or a competitor's site newly appears while yours doesn't. Read those alerts with the four tests above in mind. The alert compares the latest answer with the one before it, so it tells you something flipped, not that something real happened. Treat it as "go and look," then check whether the flip is spread across questions and whether it holds the next week.

What it doesn't cover yet: other assistants. ChatGPT is the only one connected today. For tracking, that has one useful side effect: a check that never ran is never mixed into your history as a miss, so a "drop" can't come from an assistant that wasn't actually asked.

Common Mistakes

Acting on one round. The most common and most expensive mistake. A single dip is usually chance, and a "fix" for chance changes nothing except your workload.

Asking each question only once per round. A single yes or no can't be compared between rounds, because it hides how often the yes happens. Five asks give you a rate. One ask gives you a coin toss.

Retiring a question without starting a new history. Sometimes a question really does need better wording. That's fine, as long as the new version gets a new label and starts its own history from round one, rather than being plotted on the same line as the old one.

Not logging the sources. It's the boring column, and it's the one that explains a real drop. Skip it and you'll know something changed without any idea what.

Forgetting your own site changes. If you don't write down what you changed and when, you can't line it up against a drop later. A one-line note per round is enough.

Mixing logged-in and logged-out rounds. Your account's history and location can shape the answer. If round 3 was done logged in and round 4 in a temporary chat, any difference between them might just be your account.

Treating a model update as a personal setback. When ChatGPT changes how it answers, it changes for everyone in your category at once. Check whether competitors moved too before deciding your site is the problem.

FAQ

How do I track brand mentions in ChatGPT for free?

Keep a simple spreadsheet like the log above, pick five to ten questions, freeze their wording, and every two to four weeks ask each one five times in fresh temporary chats with Search on. Record how many answers named you, how many cited your site, and which other sites showed up as sources. That's a complete tracking method with no tool at all.

How often should I check what ChatGPT says about my brand?

Every two to four weeks by hand, plus an extra round after any deliberate change to your site or listings. Checking every few days mostly measures random variation, because ChatGPT's answers don't usually shift for real that quickly.

Is a drop from 4 out of 5 to 2 out of 5 a real drop?

On its own, usually not. If ChatGPT names you about three times in five, you'll see 2 or fewer roughly one round in three purely by chance. It becomes worth acting on when the same drop shows up across several questions at once, or holds for two or three rounds in a row.

How many times should I ask the same question?

For research-grade numbers, SparkToro suggests 60 to 100 runs per prompt. For a hand-run routine, five asks per question per round is a practical minimum. It won't give a precise percentage, but it separates "almost always" from "sometimes" from "rarely," which is usually what you need to know.

ChatGPT stopped mentioning my business. What should I do first?

Rule out the causes that need no fixing. Check whether the drop is spread across several questions, whether it held for more than one round, whether your own routine changed, and whether ChatGPT's release notes show an update in between. If it's real, the sources column usually points to the reason: a new site ChatGPT is leaning on, or a change on your own site.

Should I track competitors too?

It helps, mostly through the sources column. You don't need a separate routine: when you log the sources for your own questions, note which competitors' sites appear. A competitor that starts showing up in the sources across several questions at once is usually the clearest explanation for a real drop of your own.

Does AI Schema Gen track ChatGPT mentions?

Yes, for ChatGPT today. AI Visibility tracking on Pro plans and above runs your questions weekly, records whether your site was cited and which other sites were, and alerts you when that changes. It tracks citations of your site rather than your name in the answer text, and the other AI assistants it lists aren't connected yet.

If the drop you're seeing turns out to be on your side, the fastest way to find out what changed is a free AI readiness check, which shows whether AI crawlers can still reach and read your pages and what's holding them back.