How to measure AI visibility

Start with the decision, not a score

Choose the business question first. For a software team, that might be whether a product is suggested for a specific workflow or described with accurate capabilities. For an agency, it might be whether a client’s service pages support the explanations buyers need. Choose prompts that can inform that decision.

Use sales questions or approved first-party research when available. If you write the prompts yourself, label them analyst-authored. A plausible question does not prove demand or search volume. Agree the panel before collecting answers so disappointing results do not quietly disappear.

Keep three question groups separate

• Branded understanding: “What does ExampleCo offer?” The prompt supplies the name, so a mention is weak evidence of discovery.
• Unbranded discovery: “Which tools support approval workflows for a small finance team?” Do not seed your preferred business into the question.
• Informational sources: “How should a finance team document approval exceptions?” Inspect which sources support the explanation.

These examples are illustrative, not a real experiment. Record a prompt ID and version for each. Freeze the wording and explain any changes between rounds.

Repeat tests without pooling incompatible conditions

Plan a repetition count you can review carefully and run each prompt in a clean session. Record timestamp, product, UI or API, model if disclosed, search mode, language and available region settings. Keep unknown settings explicitly unknown. Personalization and product changes can make two apparently identical runs differ.

Compare like with like. A paid consumer subscription is not simply an API model name. A search-enabled answer can consult sources that a non-search answer cannot. Report each group separately rather than pooling them into one average.

Review the answer, then the source

A mention is a name in the response. A recommendation explicitly suggests the business for the need. A citation links to evidence. One response may contain all three, just one, or none. Inspect whether a linked page supports the claim and whether the business details are accurate.

Do not call the first business in a list the “winner” unless the response explicitly ranks it and your protocol explains how that evidence is classified. A citation to a critical review is not an endorsement. Save the complete answer and cited URLs so a second reviewer can check the classification.

Use counts people can audit

Show attempted runs, successful eligible runs, failures and exclusions. An illustrative result could say “4 of 15 successful unbranded responses mentioned the business; 3 further attempts failed.” This is clearer than a percentage that hides the small sample and the failed attempts. No Takusen benchmark results are implied here.

If no tests were run, report “not measured.” Do not enter zero. Separate factual-accuracy checks from visibility: being frequently mentioned with incorrect details is a different problem from not appearing.

Connect observations to a testable change

Suppose sampled responses confuse two products. First inspect the linked sources and your public descriptions. A useful action might be clarifying a comparison page and checking that the correction is in served HTML. Repeat the same panel later, but do not attribute all movement to your edit. Models, retrieval sources and competitors can change too.

Keep referrals and inquiries separate from synthetic answer tests. OpenAI’s publisher FAQ describes ChatGPT referral attribution; it is not a census of every mention. Google’s AI features documentation describes eligibility and reporting for its own search experiences. Neither turns a small external prompt panel into total market visibility.

Keep the SEO foundations

Google’s optimization guide emphasizes useful content and established search foundations for its generative search features. Our recommendation is to prioritize clear, accurate pages and inspectable evidence over unsupported universal scores.

Use the Takusen protocol for definitions and read the crawler-access guide before interpreting missing citations as a content-quality problem.