ChatGPT Brand Tracker: How to Monitor Brand Mentions in ChatGPT

What a ChatGPT brand tracker actually tracks

A ChatGPT brand tracker asks ChatGPT the questions your buyers ask, stores the answers, and counts two things in them: whether your brand was named, and whether a page of yours was cited as a source.

That is the whole measurable surface. There is no position to record, because an answer is not a list, and no impression count, because nobody outside the model’s operator sees how often anybody asked. What exists is a sample: these questions, on these dates, produced these answers.

So the unit is one answer, and the only honest figure is a count over a stated number of answers. A tracker handing you one number with no denominator has decided on your behalf what it divided by.

Mentioned is not cited, and the gap between them is the job

These two words get used interchangeably across this category, and they describe different events with different causes and different fixes.

Mentioned means your brand name appears in the answer text. ChatGPT can do that from memory: the name sat in its training data, it is associated with your category, and out it comes when the category is asked about. Nothing of yours was fetched.

Cited means the answer carried a source and the source was a page of yours. For that, the model has to have searched and used what it retrieved. A citation is evidence that a specific URL was read.

A brand can be named from memory without a single page of its site ever being fetched.

The two come apart in both directions, which is why one number cannot stand for both. A well-known brand can be named in nine answers out of ten and cited in none, because nothing was looked up. A smaller one can be cited without being recommended.

They also demand different work. Being unnamed is a memory problem, and the route into it is being written about by other people. Being uncited is a retrieval problem: the page exists and the model used something else.

Why there is no rank to track inside a ChatGPT answer

Search gave everyone the same ten results in a fixed order, so a position was a real property of a page. An answer has none of that structure.

Ask for the best options in a category and you get prose: two or three companies named in a sentence, one often described at length and the rest in passing. Reorder the words and the set changes. There is no third slot to occupy.

Order inside one answer is not a rank either: a name appearing first can be an alphabetical accident or the company the model had most text about. So a tool offering a ChatGPT rank has manufactured it, usually by averaging where your name landed in the wording. That conversion is where the information goes.

What ChatGPT exposes to a tracker, and what it does not

Everything a monitoring tool knows comes from asking questions through an API and reading the replies. There is no feed of what other people asked.

Available: the full text of an answer to a question you chose, on a date you can record; whether the model searched while answering; and where it did, the sources it used.

Not available, at any price: how many real people asked anything; what they were told; the share of conversations in your category; anything inside the consumer apps, shopping surfaces included.

That second list is where the impressive metrics come from. A prompt-volume figure for an assistant is a model rather than a measurement: nobody publishes the underlying number.

How to track brand mentions in ChatGPT by hand

Before buying a ChatGPT mentions monitoring tool, run the measurement yourself once. It takes an afternoon and tells you whether anything here is worth watching.

1. Write ten buying questions. The ones a customer types: best tool for this job, alternatives to the incumbent, what a team of four should pick. Not what do you think of my company, which nobody asks.
2. Ask each one more than once. Three times on separate days beats three times in a minute: part of what varies is the model, part is what was on the web that morning.
3. Record the whole reply, not a verdict. The exact question, the date, whether the answer searched, and the complete text.
4. Keep two columns, never one. Named, and cited — and leave the citation cell empty rather than zero on any run with no evidence of searching.
5. Count at the end. Named in how many of how many; cited in how many of the runs that looked. Two denominators, no third number blending them.

Six weeks of that is a baseline. If the counts never move, no dashboard would have told you faster.

One answer is one observation, not a measurement

ChatGPT does not answer the same question the same way twice, and that fact invalidates most of what gets screenshotted in this category.

Sampling makes the wording vary; retrieval makes the substance vary, because an answer that searched read whatever was there that hour. So one check is a presence check: on this date, asked this way, you were named or you were not. That is a fact, and not a rate.

Repeats are what turn a presence check into a measurement, and repeats cost money, which is why every tool here decides how many it buys. Ask how many. Where the answer is one, it has to be labelled as one.

Not measured is not zero

The most damaging number in a ChatGPT brand monitoring dashboard is a zero in a cell nobody checked.

Three cases look identical on screen and are not the same thing. A measured zero: the question was asked, the model searched, no page of yours was used. An unrun check: the call failed or a key was missing. A run never asked to look: the model answered from memory, so no citation could have existed.

Only the first carries information. Printing 0 in the other two manufactures a finding: it says you were searched for and not found, when nothing looked.

The same error has a subtler form. If a brand was named in eight answers and only three of those runs searched, cited in 1 of 8 understates citation by the five runs where a citation was impossible. The correct rendering names the denominator it actually had — cited in one of the three runs that searched, and silent about the rest.

What a ChatGPT visibility tracker can honestly report

Four outputs, each a count with a stated denominator. The fourth is the one most tools skip.

None of those is a grade, an index or a single headline figure. An index needs weights, and nobody can justify the weights.

How this product asks ChatGPT, plan by plan

So the limits above are visible rather than implied, here is what each plan asks of ChatGPT.

The free check asks up to five questions, written from your own homepage, on two assistants: ChatGPT and Gemini. Each is asked once, so a result is one run per question and is labelled as one run rather than averaged. Three scans are allowed in any rolling 30 days.

On the free check, ChatGPT answers from memory. The search tool is not attached to it, so a free result can tell you whether ChatGPT named you and cannot tell you whether ChatGPT read your site: it was never asked to look. That cell reports as not measured, carrying its reason, and never as a zero. Gemini is the assistant that fetches on the free check, so the citation side is measured there under its own denominator, and the two are never added together.

On Starter and Pro, ChatGPT fetches. It is asked one day in seven and searches every one of those occasions — about four a month, each with its own date — which is what makes a ChatGPT citation measurable at all. The repetitions arrive as separate days rather than extra passes inside one scan.

That asymmetry is a cost decision rather than a feature ladder: a searching ChatGPT call is the dearest call in the system. On every plan the two counts are printed with their own denominators and never combined.

Choosing a ChatGPT mentions monitoring tool

Six questions. A vendor who cannot answer them from their own documentation is selling a dashboard rather than a measurement.

1. How many times is each question asked, and over what period? One is a presence check, and saying so is the only acceptable answer at one.
2. Does ChatGPT search on every run, or some? If some, which, and what is the citation figure divided by?
3. What appears in a cell that could not be measured? If the answer is a zero, stop.
4. Are mentions and citations ever shown over the same denominator? They have different ones. A single blended rate is the tell.
5. Where do the questions come from, and can I change them? A fixed list measures the vendor’s sample, not your market.
6. How long is the answer text kept? A citation you cannot re-read is a claim, not evidence.

None of those asks how many engines are covered, which is what this category competes on and the least useful axis for a buyer to pick on.

What ChatGPT brand monitoring cannot tell you

Four things, and all four get quietly implied by pricing pages.

Your share of what ChatGPT says to the world. The sample is the questions asked, so a percentage sold as your share of AI answers is a share of somebody’s question list.

Why you were not named. The model does not explain itself, and a reason given for an absence is a guess in a confident font.

What it earned you. A name spoken inside an answer carries no referrer, and a followed link arrives looking direct.

That a change will move an answer. The evidence is editable and the model is not, so the only honest commitment is about the work and the measurement.

Reading the numbers without fooling yourself

Read the denominator before the number. Forty per cent means nothing until you know whether it is two of five or four hundred of a thousand, and whether all thousand runs were asked to search.

Treat a small week-on-week move as noise unless the run count behind it is large. A sparkline redrawn off two observations is the variance, called a trend.

Separate the questions you could win from the ones you cannot. Being unnamed on a broad best-tools question against eleven incumbents is a different problem from being unnamed on one describing your niche. The second is fixable this quarter.

A ChatGPT brand monitoring routine you can keep

Monthly is the right reading interval, because what moves these answers is published pages.

1. Ask the same questions, word for word. A reworded question is a new question, and its answers do not belong in the old column.
2. Record named and cited separately, with the run count beside each. That count is what people drop first.
3. For anything newly cited, open the page that was used — that is which of your pages is doing the work.
4. For anything newly missing, look at who was named instead and what page of theirs was cited. A competitor’s cited page is a brief.
5. Change one thing, then wait two cycles. A fix with no measurement either side is a story about a fix.

Twelve monthly readings is a year of evidence about a surface most competitors guess at.

FAQs