1. Define the observation
Choose what you want to observe before running prompts. For example, you might ask whether an assistant links to a particular public guide when answering a defined set of customer questions. Record a mention and a link separately. Distinguish an accurate description from a recommendation, and a recommendation from a measured referral. The question set should represent the decisions you care about, including some questions that do not mention your brand. Keep the initial set fixed for comparisons. If you revise it, create a new version rather than quietly replacing difficult questions with ones that produce flattering answers.
2. Preserve the conditions and denominator
For every observation, save the exact question, product or model shown, date, language, relevant settings, answer, and cited URLs. Note unavailable information rather than guessing. Use a fresh conversation when that is part of your protocol, and document any follow-up prompts. If you summarize results, state both the count and denominator: observed linked citations out of the completed runs in this particular sample. Explain missing or failed runs separately. This is a sample description, not a market-wide visibility percentage. Repeat under comparable conditions, but do not claim that a repeatable process makes variable answers deterministic.
3. Understand what the tools establish
Google’s current guide applies foundational SEO to generative AI search; eligibility does not guarantee appearance. Search Console also documents a dedicated generative AI impressions report. Read its coverage and availability before drawing conclusions; an overall Web total is not an isolated AI metric. Suprara’s scanner provides a separate layer: public HTML, robots.txt and sitemap checks. It does not retrieve that report or verify that a particular property has data. Keep implementation evidence, provider-reported exposure, manual answers and business outcomes in separate report rows.
4. Report observations without causal shortcuts
Use a report with three parts: what was tested, what was observed, and what remains unknown. If more citations appear after a content update, describe the timing without asserting that the edit caused the change. Other inputs, product behavior, and sample composition may differ. Where authorized referral or conversion data exists, review it separately and document attribution limits. Improve content because it answers a real question with clear, supportable information. A citation can be useful evidence of one response; it is never a promise of future placement. Honest measurement gives your team a basis for decisions without selling certainty it does not possess.
What to take away
- Count mentions, citations, and visits separately.
- Publish the sample and testing conditions with any summary.
- An audit score is not an AI visibility measurement.
Apply what you learned
Practice this topic in Academy, then use the worksheet to document the evidence and the change you plan to make.
Sources and further reading
A clearer website starts with evidence.
Get a real public-page scan and a personalized PDF, then use the free worksheet to prioritize your next changes.