How we measure AI search visibility

1. Agree on a fixed set of buyer questions

Between 5 and 30 questions your buyers actually ask — “who repairs industrial compressors near Denver”, “is it safe to buy game keys from X”, “best personal injury lawyer in Pensacola”. They are built from demand data (search volumes, People Also Ask, your own inquiries), not from an agency's keyword list. For an engagement the set is frozen, so that month two is comparable with month one.

2. Repeat each question

Each question is run at least twice per engine: Google AI Overviews, ChatGPT search, Perplexity, Gemini, and others on request. If the two runs disagree, a third decides. A mention or a citation counts only if it repeats. Everything that does not repeat is kept in the raw file but does not enter the report.

3. Record brand mentions and source citations

Is your business named in the text of the answer, and is your site cited as a source — yes or no, per question, per engine. These are two different findings: an engine often names a company without linking to it, and the domain-only check would wrongly report “absent”. If the answer is no: who is named instead, and from which pages the engine assembled the answer — that is the content plan.

4. Check the website and search positions

• Positions in Google for the same questions, dated, same location each time.
• Whether search and retrieval crawlers (such as OAI-SearchBot, PerplexityBot and Googlebot) receive the same page as a human — robots.txt, noindex, JavaScript-only rendering, hosting firewalls that block bots.
• Structured data: whether the site tells machines what the page says — organization, product, FAQ, article, breadcrumbs — and whether it matches what the page shows.

5. Compare the same questions before and after

The “before” snapshot is taken on the day work starts, before any change. A month after implementation the same set is run again, one to one. The result is a table: which questions gained a mention or a citation, which did not, and what we think is the reason. A platform-wide shift (an engine changed how it answers) is named as such, not booked as a win or a loss.

Scope and limitations of the measurement

It is not a share-of-voice score, not a “visibility index” and not a promise of a position. It is deliberately conservative: it undercounts rather than overcounts, so that a gain reported to you is a real one. The raw results are yours. Anyone can repeat the recorded method, although the engine may return a different answer later.