How to track brand mentions in ChatGPT (and why one screenshot is not data)

You cannot install a tracking tag inside ChatGPT. There is no referrer for an answer the model composed from its own weights, and no console OpenAI hands you showing how often your brand came up this week. So tracking brand mentions in ChatGPT is not monitoring in the analytics sense. It is sampling. You decide which questions matter, you ask them through the API on a fixed schedule, and you count what comes back.

A method that survives scrutiny has four parts.

• A fixed prompt set. The questions your buyers actually ask, written once and then left alone, so a change in the trend means the market moved and not that you rewrote the questions.
• A schedule. The same set, run at the same cadence, so this week is comparable with last week.
• An extraction step. Every answer read for three things: whether your brand is named, where it sits in the answer, and in what tone.
• A denominator. Your mentions set against mentions of the competitors you actually lose deals to. A bare count cannot move in a way you can interpret.

The rest of this guide is the detail of those four parts, in the order a learner needs them, plus the two places this measurement is easy to overclaim: the share figure and week over week movement.

The usual first move is to open ChatGPT, type "best tools for X", and screenshot the result. It is a reasonable instinct and a bad measurement. Five reasons, and each one is also a design requirement for anything better.

• The same prompt returns different answers. These models sample from a distribution rather than looking up a row. Ask twice and the wording changes, and often the list of brands changes with it. One reading gives you one draw from that distribution and no sense of its spread.
• Phrasing moves the answer more than you expect. "Best tools for mid-market operations teams" and "top workflow automation vendors" are the same question to a marketer and different questions to a model. If you vary the wording each time you check, you are measuring your own phrasing.
• Your account is not a neutral observer. In the consumer app, memory, custom instructions, prior turns in the thread and your own history all shape the answer. Searching for your own brand from an account that has been researching your own brand for a year is the least representative sample available.
• The model underneath changes without warning. Providers ship new versions. An answer set read on one version and compared to one read on another is not a clean comparison, and if you have not recorded which version answered, you cannot tell which kind of change you are looking at.
• A single reading has no baseline. Absent from one answer tells you nothing. Absent from most answers to the twenty questions your buyers ask, four weeks running, is a finding.

None of this makes AI answers unmeasurable. It makes a single reading unusable and a sampled series useful. The honest claim from sampling is directional: across this fixed set of questions, over this window, your brand was named in roughly this proportion of answers, and here is the caveat.

Your prompt set is your measurement instrument. If it is skewed, nothing downstream can correct it. A usable set covers several distinct kinds of question.

• Discovery. "Best tools for [job to be done]", "who are the leading vendors in [category]". These are where absence hurts most, because the buyer has no brand in mind yet.
• Qualification. "Best [category] for a 200 person B2B company", "which [category] tool suits a regulated industry". Segment-shaped questions reveal whether the model associates you with your actual buyer.
• Comparison. "[You] vs [competitor]", "[competitor] alternatives". Here you are usually mentioned, so the interesting variable is how you are characterised and in what order.
• Attribute. "Cheapest", "easiest to implement", "best for agencies". These expose the adjective the model attaches to you, which is often the thing you most want to change.
• Branded control. A handful of prompts that name you directly. They will not move much, which is exactly the point: they are your check that a sudden drop elsewhere is not a collection failure.

Two practical notes. First, write prompts in the language and market you sell into, because answers differ by both, and do not present a single-language sample as global. Second, resist stuffing the set with prompts you already win. A set built to flatter produces a number that cannot fall, which is the same failure as having no competitors in the denominator.

The extraction step is where most homemade tracking quietly breaks, because a human skim is not reproducible. Decide the fields once and apply them identically to every response.

• Mentioned or absent. Binary, on the brand name and its obvious variants. Absent is a data point, not a gap in the data.
• Ranking position within the answer. Being named third in a list of eight is a different commercial fact from leading it, and position moves earlier than presence does.
• Sentiment, in coarse buckets. Positive, neutral, negative, absent is enough. Be sceptical of any sentiment label, including your own tooling's: automated classification reads keywords, and keywords mishandle negation and hedging. Treat sentiment as a flag that tells you which answer to open, not as a verdict.
• Which competitors were named. This is your denominator and also your most useful qualitative signal.
• Provider, model version and timestamp. Non-negotiable, for the reasons in the drift section below.
• The answer text itself. Stored as evidence, with personal names, emails and handles redacted before storage. Model answers do sometimes name real individuals, and keeping raw text you do not need is an obligation you do not need either.

If you build this yourself, the sentiment classifier will be the part you rewrite most often. Word boundaries matter, negation matters, and an unbounded keyword match will confidently tell you that "unlimited" is negative.

Counts are hard to argue with and hard to act on. What a board deck needs is a share: of the brand mentions across your prompt set this week, what proportion were yours.

The arithmetic is simple. Take your mentions across all answers in the window, divide by your mentions plus the mentions of your named competitors, and express it as a percentage. Answers where nobody was named still belong in your records as absent, because a category question that produces no vendor at all is a finding about the prompt, not a missing observation.

Three rules keep the figure honest.

The denominator must contain rivals. With no named competitors, the formula returns a permanent perfect score. That is the single easiest way to turn a measurement product into a flattering one, and it is worth checking in any tool you evaluate.

An unsampled week is not a zero. If the run did not complete, report no score. A large zero under a headline saying there is no score to report is worse than a blank.

The caveat lives on the record. Store the methodology note with each computed score so it appears wherever the score appears. Week over week movement in an LLM-sampled metric can reflect a model update rather than any change in how the market sees you. A number that travels without that sentence will eventually be used as though it were a rank in a search engine, which it is not.

There is no published benchmark for a good share of AI voice, and anyone offering you one is guessing. The useful reading is relative: your share against the three names you compete with, and its direction over thirty days.

This is the question that decides whether your tracking is trusted internally. Your share falls from thirty-four per cent to twenty-six. Did competitors gain ground, or did the provider ship a new model that answers your category differently?

You cannot answer that from the numbers alone. You can answer it from metadata, which is why every stored answer needs its model version. When the version string changes between runs, treat the first week on the new version as a re-baselining event rather than a result. Annotate it. Say plainly in the report that the comparison spans a model change and that the movement is not attributable yet. Then watch the following two or three weeks on the stable version before drawing a conclusion.

The same discipline applies to your own changes. If you added prompts, changed a competitor name, or switched providers, the series before and after are different series. Note the break rather than smoothing it over.

This is also where your branded control prompts earn their place. If presence on branded prompts held steady while category prompts dropped, something changed in how the model reasons about the category. If branded prompts dropped too, look first at collection.

Being specific about the limits is what makes the rest defensible.

It is not traffic and it is not attribution. A mention is not a click, and you will not learn from sampling how many buyers acted on an answer. Treat share of AI voice as a presence metric that sits beside your analytics, not inside them.

It is not what every user sees. An API sample is unpersonalised by design. Real users bring memory, custom instructions, prior turns and their own history. Your sample is a controlled read of the model's default disposition toward your category, which is the most comparable thing available and still not a census.

It does not cover every assistant. Two providers is two providers. If a tool implies it watches all of AI search, ask which endpoints it actually calls and how often.

Nothing here is a ranking you can optimise deterministically. There is no documented ordering function to game, and any promise of a guaranteed position in an AI answer should be read as sales copy. What you can do is measure, find the prompts where you are absent, fix the underlying evidence on the open web, and measure again.

Provider access is a dependency, not a given. Rate limits and usage policies change. Any setup you rely on, built or bought, should be designed to fall back to the remaining providers rather than stop, and should tell you which providers a given week's number came from.

Once the collection works, reporting is a short, repeatable act. Five lines, every week, in this order.

One figure, one caveat

Your share of AI voice for the window, the direction over thirty days, and the sentence explaining that LLM answers are non-deterministic and that some movement reflects model updates. The caveat goes next to the number, not in an appendix.

The three names beside you

The same figure for each named competitor from the same prompts. This is the comparison your executives will make anyway, so make it correctly.

Where you were absent

The prompts that produced an answer with no mention of you. This is the only part of the report that generates work, and it is the part worth arguing over. Absence on a high-intent qualification prompt matters more than absence on a broad one.

One answer, quoted

A single response, in the model's own words, describing your category or a competitor. Nothing moves a content roadmap faster than reading the recommendation your buyer just read.

What changed in the instrument

New model versions, failed runs, prompts added. Boring, and the reason anyone believes line one.

If you are starting this week, do the unglamorous part first: write the twenty questions your buyers ask, agree the three competitors you are measured against, and decide the cadence you can actually keep. The counting is the easy half. Choosing a fixed set of questions and refusing to change it is what turns a screenshot into evidence.