ChatGPTStart with a question set you can explain
A useful baseline begins with the buyer task, not a headline score. Save the exact wording, category, market and approved brand aliases for each question. Keep questions that name the brand separate from broader discovery questions.
Use the buyer question selection guide to build a small, reviewable set before expanding it. Changing the set changes what a later percentage describes.
Inspect the observation behind the result
A collected observation keeps its original question, observed surface, collection date and answer evidence. Current adapter and integration statuses are published in platform readiness; production acceptance is pending. A model API response and a consumer website answer are separate surfaces.
Open the answer before interpreting a mention. A recommendation, passing reference and quoted comparison can all contain a brand name with different meaning. Inspect the evidence rather than treating every name match as a recommendation.
Keep coverage separate from mention rate
Count eligible answers before calculating a rate. A failed or missing answer is not a valid answer without a brand mention.
| Measure | Counting unit |
|---|---|
| Coverage | Valid answers / planned answer runs |
| Mention rate | Valid answers naming the brand / valid answers |
| Owned citation rate | Answers citing an owned page / answers with complete citation evidence |
These are explicit measurement conventions, not universal vendor definitions. See the measurement walkthrough for an illustrative calculation.
Compare a later sample with the same context
Keep a record of edits to prompts, brand aliases, competitors, market and collection method. Describe a changed baseline rather than presenting it as the same experiment.
A later sample can show a movement. It cannot establish that one page edit caused the movement. Connect the observation to an assigned action and record what was actually changed, then review the same approved question set.
Bring evidence into the client conversation
A client review should show the sample population, unresolved coverage, source URLs and the status of the work. Lead with the decision the evidence supports and the limits it does not resolve.
Use the AI visibility report template to keep denominators, findings and follow-up work together. If the sample is too small or coverage changed, report that directly instead of hiding it inside an overall score.