Design for the question behind the query
A useful answer states scope, definitions, assumptions and evidence directly. Retrieval-query traces can reveal which qualifiers and subquestions matter in practice.
Keep evidence provenance visible
Distinguish what a provider surfaced from what a model generated as a plausible alternative. Mixing those classes creates false confidence and weak optimization decisions.
Validate outside the model
Measure citations, referrals, task completion and business outcomes. Use query estimates as exploration—not as synthetic search volume.