Start with a commercial question, not a vanity keyword list
Choose questions that describe a real purchase or evaluation decision: best software for a particular team, alternatives to a known product, or products that satisfy a concrete constraint. Record the wording, locale, region, and provider surface because each can change the answer.
A long prompt inventory can create impressive charts while hiding the questions that matter. Begin with a small, confirmed topic set and add questions only when someone can explain the business decision they represent.
Repeat and preserve the native evidence
One answer is an observation, not a stable market fact. Repeat the prompt, use more than one compliant provider where possible, and preserve raw response fingerprints plus the citations returned by that surface. Do not ask a second model to manufacture URLs that were absent from the native response.
Separate value from confidence
Opportunity Score estimates the value and actionability of closing a gap. Evidence Confidence measures recurrence, source quality, freshness, and agreement. A valuable question can still have low confidence; the correct next step may be more sampling rather than a large content project.
Recheck under comparable conditions
After an approved action is completed, reuse the stored prompt, provider, locale, region, and sampling plan. Label the outcome Won, Improved, No change, or Insufficient evidence. The result can support a correlation, but it cannot prove that the action alone caused a generative system to change.