If you monitor how ChatGPT presents your business, you need a repeatable record of answers to questions your buyers ask. Set up a prompt set, send it to ChatGPT on the days you pick, and read what comes back over weeks. Not as a single check.
Build the prompt set from real questions
Start with what your buyers ask, in their words. Sales calls are the best source you have. Whatever a prospect asks in the first ten minutes is what they would have typed.
Cover category questions, comparisons between named products, problem questions that never name a category at all, and questions about your brand by name. Most teams forget the last group, and it is where the fixable problems live, because an answer that describes your product wrongly got that description from somewhere you can find.
Thirty prompts you chose deliberately beat three hundred generated off a keyword list. How to choose the prompts you track covers the selection.
Choose the days, and understand the timing
Tick the days you want fresh data on. Collection runs the night before each ticked day, so a Monday tick is a Sunday night run.
A prompt takes appreciably longer to collect than a search. The run starts the night before for that reason, not at breakfast. If you are used to hitting refresh on a rank tracker and watching it fill in, this does not work like that. The report is ready for the day you asked for.
Cost is a multiplication: prompts, times providers, times days ticked. A hundred prompts on ChatGPT alone, twice a week, is 200 calls. Across all three providers it is 600.
What is captured from each answer
The collected result carries the answer text, the sources it showed, the web results it used, the follow-up queries it reported, and the brand entities it identified.
Nothing goes back over the answer afterwards looking for brand names. What a report shows for a run is what came back at that moment, not a later interpretation of it.
The follow-up queries are what the service said it would search. They tell you how a question decomposes. They are not proof of what ran inside the service, and nothing observable from outside would be.
Reading the result over time
One run is one observation of a system that answers the same question differently on different days. Do not act on it.
Look for consistent presence, consistent absence, direction across weeks, and how you compare against competitors facing the same conditions. Why repeated answers vary covers the mechanism.
Read a month, not a morning. Named in seven runs out of ten across four weeks means you are established for that question. Named in none of them is a real problem you can do something about.
Turning the result into work
What you find lands in one of these, and each needs different work.
Named inaccurately. Read the answer and go find where the description came from. It is usually an out-of-date third-party page and not your own site, and a correction request to that publisher is the shortest route there is.
Absent while competitors appear. Look at the sources those answers cite. Those pages are being used to answer your buyers, so whether you are on them is a placement job with the targets already named.
Cited but not named. Your page is a useful source and the answers built on it recommend somebody else. Check whether the cited page says what you sell, or only what you know.
What this does not tell you
It does not attribute traffic or revenue. The path from an answer to a visit is indirect when it exists at all. It does not explain why ChatGPT picked its sources, which nobody outside the service can see. And it guarantees nothing, because nobody controls the thing being measured.
At your visibility review, return to the tracked questions and read the answers behind any change. You can then choose a specific issue for further research.
See AI rank tracking, the same work in Gemini, or how collection runs.