Start with the problem the user will ask
Branded prompts verify whether a model recognizes a known name; they do not measure spontaneous discovery. Generic-problem prompts are stronger controls because they state a job without telling the provider which product to find.
Keep a fixed control set and add bounded variants. If the prompt changes every day, a positive result cannot be compared with yesterday's absence. Record prompt fingerprints, provider, model, and observation time.
Separate mention, recommendation, and support
A mention places the product name in an answer. A recommendation connects it to the requested job with affirmative language. Evidence support identifies which consulted URLs contained enough product information to make that connection inspectable.
Owned documentation can support a correct recommendation, but it is not independent endorsement. Label the evidence as owned source, directory record, independent source, or mixed. Keep a single-provider result single-provider until another provider repeats it.
- Quote the exact recommendation excerpt.
- Link the first product source in the search trace.
- Separate later branded verification from the original unbranded entry query.
Amplify a right path without copying noise
When an unbranded query leads to a product source and an explicit recommendation, preserve that route as a right-path receipt. Build deeper pages for adjacent intents, improve the facts on sources already consulted, and rerun the original prompt unchanged.
Do not turn one result into a promise. The useful claim is that a stated provider produced a recommendation under stated conditions on a stated date. Replication is the next experiment.