Define the decision and sample
Start with a decision such as which customer questions deserve better documentation. Select a small set of real questions before collecting responses, and preserve the exact wording. Include the language and audience context that actually matter to your business. Document how you selected the questions. This is an editorial research protocol, not a representative survey of every possible AI interaction. Its value comes from a clear boundary and repeatable collection.
Build a record that can be checked
For each run, save the provider, product or mode, timestamp, prompt, full response, cited URLs, and relevant settings. Add distinct fields for entity mention, citation to your domain, factual accuracy, and recommendation. Allow “unclear” and “request failed” instead of forcing a yes or no. A second reviewer can examine ambiguous cases using the same definitions. This does not eliminate judgment, but it makes disagreements visible rather than hiding them inside a score.
Compare like with like
Keep the original question set when repeating the study. If business needs require new questions, label a new cohort rather than silently replacing the old one. Report completed observations, failures, and any exclusions. Compare the same language and mode where possible. A change may reflect the answer system or the sample, so describe patterns as observations. Avoid a causal claim unless the research design and evidence genuinely support it.
Translate findings into a bounded action
A useful finding might be that several saved answers misstate your cancellation policy. First verify the current policy on your pages, then identify ambiguity or outdated material you can legitimately correct. Keep the study log separate from Suprara’s public-page scan and PDF. The study records actual AI outputs; the scan captures initial HTTP/HTML, robots.txt, and one sitemap within its stated scope. Use each for the question it can answer. After publishing changes, recheck the same URL and preserve both reports without presenting the scan as comprehensive AI visibility.