Start with exact identity
Product families are convenient for marketing, but they are dangerous evidence boundaries. Adjacent models can differ in dimensions, certifications, compatibility, included components, warranty terms, and lifecycle status. An accurate family-level statement can still be wrong for the exact model a buyer is evaluating.
Record the full model or SKU before evaluating a claim. If configuration changes the answer, preserve the relevant option, region, or revision. When identity remains ambiguous, classify the result as unresolved rather than transferring a fact from a neighboring product.
Preserve the buyer-facing observation
An audit needs the original AI evidence, not a paraphrase. Retain the provider, model when visible, exact query, verbatim answer, capture time, citations, and a screenshot or raw-response reference when available. Note whether web retrieval was active and any observable market or session conditions.
This record lets a reviewer distinguish an affirmative false claim from a cautious answer, a missing citation, or a recommendation gap. Those outcomes require different language and different corrective action.
Apply a source hierarchy
Compare each material claim with a current manufacturer or regulator source. Product pages, current specification sheets, manuals, warranty documents, certification records, and lifecycle notices can control different claim types. Dealer and distributor pages are useful propagation evidence, but they should not override current first-party product truth.
Record a short controlling quotation or precise source location so another reviewer can reproduce the decision. A URL alone is not enough when a long manual covers multiple models or configurations.
- Identity gate: The evidence names the exact model or SKU.
- Authority gate: The source is current and appropriate for the claim type.
- Materiality gate: The discrepancy could affect discovery, selection, use, compliance, or support.
- Review gate: A second reviewer independently reaches the same conclusion.
Separate findings from questions
A defensible audit separates verified errors, visibility gaps, source conflicts, correct answers, and unresolved questions. Missing evidence is not proof that a negative claim is true, and a single variable AI response is not automatically a repeatable market pattern.
After an approved correction is published, rerun the original buyer question under documented conditions. The goal is not to claim control over an AI platform. The goal is to measure whether the buyer-facing answer and its supporting sources improved.