An AI visibility audit you can actually repeat

Choose a bounded question set

For a first exercise, choose ten questions from actual customer conversations. Ten is a manageable starting point we suggest, not a statistically representative sample. Include learning, comparison, and verification tasks relevant to your business.

Record why each question matters and which page should help answer it. Keep brand-led questions separate from questions that do not name your brand. Changing the mix can change the apparent result even if nothing else has improved.

Keep a record a colleague could inspect

For every observation, retain the exact question, date, platform, visible settings, response, and linked sources. Include the language and whether the test was performed in a signed-in session. If a variable is unknown, label it unknown.

Repeat without moving the goalposts

Keep the same questions and observation method for the next round. Record material interface or model changes where they are visible. If you change the question set, create a new version rather than silently rewriting the baseline.

Report a small sample as a small sample. “Named in four of ten recorded answers” is more interpretable than “40% market visibility.” Repeated runs may help reveal variation; they do not turn a convenience sample into a market-wide estimate.

Turn observations into a narrow action list

Choose actions tied to specific findings: correct an outdated product explanation, make a useful source page reachable, or add missing comparison criteria. Give each action an owner and a completion check.

In parallel, inspect your own referral and conversion data using stable definitions. The observation log tells you what was visible in sampled answers. Your website data tells you about recorded visits and actions. Bringing them into one review is useful; combining them into one unexplained score is not.