The short version
We ask ten buying questions in a logged-out AI session, three times each, over three days. We keep every answer and every cited source. We then read the answers by hand, count mentions and recommendations, and classify the sources. The report shows the three runs separately; it does not collapse them into one “rank”.
Why three runs
Ask the same question twice and you can get different names in a different order. A single answer is a screenshot of one moment. Three runs let us show whether a brand is consistently absent, consistently present, or moving. The report keeps the runs visible because the spread is the finding: a brand that appears once in three answers has a different problem from one that appears in all three at position six.
The free check on the homepage is deliberately one run per question. It is a demonstration of the mechanism, not the paid measurement. The diagnostic runs three and keeps the dates.
What the report counts
- Mention rate: the share of the thirty answers in which the brand appears at all. A mention is not a recommendation; a list inclusion is not a win.
- Recommendation rate: the share of answers where the brand is put forward as a choice, not only listed as an option. This is a judgement call, and the report shows the sentence behind every count so it can be challenged.
- Position: where the brand first appears in the answer relative to the named competitors. Position is descriptive, not a universal ranking.
- Citation share: the sources the answers cited, classified by type and counted by appearance. A source appearing in several answers is counted each time; the report says so.
- Variance: the difference between runs of the same question. We report it rather than hiding it behind an average.
What the sampling looks like
The paid diagnostic uses logged-out ChatGPT sessions in one location, in US English, with no account history, extensions or memory. The question set is fixed before sampling. The three runs are spread across three days, not run back-to-back, so a transient answer is less likely to be mistaken for a stable one. The exact model surface and dates are written into the report.
The free check uses Perplexity Sonar through OpenRouter as a fast web-grounded proxy. It is a different surface from the paid ChatGPT sampling. The interface says so, and the results are not presented as representative of ChatGPT. For the monitoring version of the same questions, see ChatGPT brand monitoring and AI search metrics.
What we do with the sources
Every citation is opened and classified into one of five buckets: directories, community, editorial, first-party, or video and retail. The classification is deliberately simple; the point is to show which kind of page is doing the work, not to produce a perfect taxonomy. If an answer cites a page that is not reachable, we record the failure rather than filling the gap with a guess.
What we will not claim
- That AI answers are stable, or that a single “rank” exists across sessions and users.
- That a citation proves the cited page supports every claim in the answer.
- That a recommendation is an endorsement, or that a mention will produce revenue.
- That the work will produce a citation, a lift or a particular position by a particular date.
- That a synthetic sample is a client result. The sample reports are labelled synthetic on every page.
How to read a three-run result
Start with the questions where the brand is absent in all three runs. Those are the clearest gaps. Then look at the questions where it appears in one or two runs: the issue may be instability rather than absence, and a different intervention may be needed. Then read the cited sources for the questions where a competitor appears and you do not. The action is usually on those pages, not in a new blog post.
Read the synthetic 8-page sample · See the diagnostic · Run the free check