Key takeaways
- Retain original answers.
- Read the sources.
- Label explanations as hypotheses.
Define what you observed
Record whether the answer names the business, links to it, recommends it for a purpose or quotes a source mentioning it. These are different outcomes. A cited page does not reveal every input used to produce the answer.
Create a comparable test
Keep the task, wording, assistant mode and market consistent. Save the full answer rather than a cropped positive screenshot. Include the date and visible settings. A changed prompt is a new test, not another observation in an unchanged series.
Read the cited pages
Ask what each accessible source contributes: a directory entry, product detail, independent review or vendor-written list. Check whether the information remains accurate. Do not assume every uncited statement came from a linked page.
Turn a gap into a hypothesis
Suppose comparison answers repeatedly discuss export formats, but your product site does not explain exports. Publish an accurate export guide and measure again. Even if mentions improve, that does not prove one private ranking factor caused the change.
Use language the evidence supports
Write “appeared in six of twelve recorded answers” rather than “owns half the AI market.” Include missing answers and failures. A transparent report gives the next review a useful starting point.
References and further reading
Primary documentation and product references. Practical checklists and worked examples above are editorial guidance, not claims about private ranking algorithms.