The most important thing: Use a repeatable sample of real buying questions and connect visibility to qualified business results.
1Create a baseline question set
Start with 12–20 realistic questions across discovery, comparison, price, fit and branded verification. Record language, market, date, product/model and whether web search is enabled. Keep the wording stable for comparisons; add new questions as a separate cohort.
A documented prompt set can be rerun under comparable conditions.
2Record what actually appeared
Capture the answer and cited URLs. Mark whether the business was named, its website was cited, it was recommended and the facts were correct. Track competitors and external sources. Repeat a subset to understand variability; one answer is not a stable ranking.
Each observation has a retained answer or screenshot and a clear classification.
3Inspect first-party reporting
Use Search Console for organic performance and its Generative AI performance report for the available AI Search impression breakdowns. Check analytics referrals and qualified conversions. OpenAI describes ChatGPT referral tracking; no-click influence and privacy limits leave gaps.
A monthly report separates platform impressions, observed answers, referrals and qualified leads.
4Run controlled improvements
Choose a specific problem: wrong service area, inaccessible page, absent pricing context or missing evidence. Fix it, log the date and compare later observations with the baseline. Account for competitors, seasonality and platform changes before claiming causation.
Each experiment has a hypothesis, action, evaluation date and honest result.
A concrete example
Avoid this mistake: The planner’s completion percentage is only your checklist progress. It is not a measured AI visibility score, ranking prediction or traffic forecast.