Record the test context
Capture the prompt identifier, exact wording, date and time zone. Record the product name and any visible model or mode selection, search setting and relevant conversation context. If a detail is not exposed, mark it unknown rather than guessing. Use a defined test account and record relevant settings according to the agreed method. Do not include credentials or unnecessary personal information in the log. The purpose is to make the observation interpretable, not to claim you can see hidden product behaviour.
Preserve the answer before summarising it
Save the complete relevant response through an approved method, along with the visible source references. A short excerpt may omit a qualification, a negative statement or the fact that another business was actually recommended. In a fictional supplier comparison, your brand might appear only in a warning about a mismatch with the requested service. A tally of names alone would misrepresent that result. Keep enough evidence for another reviewer to read the surrounding wording and judge what the assistant actually said.
Check the visible references
Record the linked source title and destination where available, then inspect whether it supports the associated statement. Distinguish your own website from a third-party source mentioning your business. A source link appearing in the answer is an observation; it does not establish why the assistant selected it or how it influenced the wording. If a reference cannot be opened, note the limit. Do not replace an unavailable source with a different page and present that substitute as the source originally shown.
Use explicit result labels
Record mentions, recommendations, citations and inaccurate claims separately. Include whether the answer addressed the intended buyer question. A brand can be mentioned without being recommended, and a recommendation can lack a visible citation. Keep failed runs, refusals and incomplete answers visible rather than silently excluding them from the denominator. Define an unverified category for cases that cannot be assessed. The log should allow a reviewer to reconstruct the reported totals from the saved evidence without relying on the original tester remembering what happened.
Separate answer evidence from referral traffic
If you also inspect website analytics, keep those records in a related but distinct measurement stream. An attributed visit is evidence about that recorded visit, not proof of the exact answer shown or the frequency of recommendations. Likewise, a saved answer naming the business does not establish that a buyer clicked or enquired. Link observations only where you have evidence supporting the relationship. Avoid presenting traffic, citations and leads as interchangeable measures just because they all relate to assistant use.
Retain corrections and review decisions
If a reviewer changes a result label, keep the reason and original evidence. Store the log where authorised colleagues can access it and apply the agreed retention process. Record when a source or answer capture is missing instead of recreating it from memory. Use the evidence to support a bounded report stating which tests ran and what was observed. A single answer may identify a useful issue to investigate, but repeated testing is needed before describing a stable pattern within the chosen test set.