Fix the sample before the run
Build the query set before looking at the answers. Cover branded, category, problem, method and comparative intent. Keep the wording and target market unchanged during a measurement window, so later runs remain comparable.
Record what the system actually did
Preserve the full answer and every source URL. Mark a brand mention, source citation, accurate description and explicit recommendation as separate observations. One should never be used as a substitute for another.
Compare like with like
Hold the language, market, account state and answer surface constant. Report results by engine and query family before producing any aggregate. A blended percentage can otherwise hide both strong and weak areas.
Publish the limits
Generative answers vary, interfaces change and some products do not expose the same evidence. A responsible report states the sample size, inaccessible surfaces and collection date alongside every result.
Practical checklist
- Freeze the query set before collection
- Capture the full answer and cited URLs
- Separate mentions, citations and recommendations
- Retain locale, market, surface and date
- Compare only equivalent measurement windows