Preserve the answer before judging it.
Save the full response, exact question, collection date and AI experience. Keep visible citations with the text they support. A short excerpt can hide a qualification or warning elsewhere in the response.
Record failed requests and unavailable answer surfaces separately. A collection failure does not establish that your brand was absent from a valid answer.
Distinguish a mention from a recommendation.
A brand can appear as background context, an option in a shortlist, a qualified recommendation or an explicit preference. Label these outcomes separately. Position in a list alone does not establish that the system prefers one brand.
- Mention: the brand appears, without a reason to choose it.
- Shortlist: the brand is included among relevant options.
- Recommendation: the answer connects the brand to the buyer’s needs.
- Explicit preference: the answer states a reason to choose the brand over alternatives.
Check facts against a current source.
Extract each factual claim that could affect a buying decision: product dimensions, service coverage, delivery terms, return rules or price. Compare it with an authoritative, current source and save the source URL and review date.
Mark claims as supported, contradicted or unresolved. Keep subjective judgments separate from factual errors. “Too expensive” is an opinion; an incorrect stated price is a claim you can check.
Treat citations as evidence to inspect.
Open the cited page and check whether it supports the associated claim. Record the publisher, page URL and relevant passage. If no source is visible, record that absence rather than inferring a source.
A visible citation tells you what was linked in that response. It does not reveal every influence on the answer. Keep the observed evidence and your interpretation distinct.
Choose an action with a clear owner.
Group findings by the buyer decision they affect. Prioritise material factual errors and unresolved buying questions. Identify the page, content asset or source that your team can improve, assign an owner and define what a completed change looks like.
Revisit the same questions under recorded conditions after the work. A later change in AI answers is an observation to investigate; it does not by itself prove that your edit caused the change.