What a ChatGPT brand tracker should record
A tracker is only as good as its record. For each buyer question, write down five numbers on the same row:
- Mentioned — did the answer name the brand at all?
- Position — if it named a list, where did the brand sit?
- Site cited — did the answer link to the brand's own domain, or to a third-party page?
- Competitors named — who else appeared in the same answer?
- Date and engine — when was the answer collected, and from which model.
Without the date and the engine, the row is not evidence. A tracker without dates is a screenshot folder.
Why one answer per question is not a measurement
We re-asked the same 15 buying questions three times on the same day. The set of brands named in an answer was identical in only 6 of 15 questions; the average overlap was a Jaccard of 0.843. The rest moved between runs. If a tool reports a single run, its “current visibility” number is noise — and any week-to-week change it reports may be sampling variance rather than progress.
The published study, the method and the raw CSV: 45-answer variance pilot.
How to run a free one-run check (and what it can't tell you)
Start with the free check on this site: it asks three buyer questions once each, records whether the brand was mentioned and which sources the answer cited. It is enough to answer one question — is the mechanism working at all? — and not enough to answer another: is the gap stable?
If the free run shows the brand missing or uncited, the next step is the same three questions on three different days. That is the minimum unit that turns a sample into a trend.
What to do with the rows
Group the rows by question, not by brand. For each question, list the pages that won the citation. Those pages are the buyer's evidence, according to the engine. The content decision is straightforward: publish or improve the page that answers the question the same way, and re-run the question after the change.
For the pattern across 35 brands and 105 answers, see the AI visibility index. For how the measurement works, see the method note.
Common tracker mistakes
- Recording only the brand. If the answer cites a comparison page, the competitor's page is the finding.
- No dates. An AI answer is a dated artifact. Undated rows cannot be compared.
- One run. It cannot separate movement from variance.
- Running only the money question. The category question and the comparison question usually behave differently.