Start with the claim.

Take one specific statement from the answer: for example, that a running jacket is reflective. Inspect the cited page and record whether it supports that statement, qualifies it or says nothing about it. Keep the original response and the page evidence together.

Separate overlap from influence.

A product page and an AI answer may describe the same features. That correspondence is useful evidence, but it does not establish that the page caused the wording or recommendation. Other sources, prior model knowledge and the testing context may also matter.

Use plain alignment labels.

Strong alignment means the page supports the relevant claims being checked. Partial alignment means it supports some of them. Context only means the page helps explain the topic without substantiating those product claims. These are descriptive labels, not an algorithmic visibility score.

Keep missing evidence visible.

If no sources are supplied, report that. If a page cannot be accessed or changes later, record the limitation. Do not fill an evidence gap with a plausible story.

Turn the observation into a test.

A missing or inconsistent product fact may suggest a useful change. Verify the fact first, record the intervention, then rerun the same benchmark. Include unchanged and worse results. A before/after difference alone still does not establish causation.

This is a method note, not a completed experiment or a claim about any particular AI provider.