A business name in an AI answer is an observation worth keeping. Before turning it into a visibility score, decide which questions were asked and what counts as a mention.
Keep the demonstration in its own frame
Julian Goldie reports nine mentions across 40 answers in his Search Agent OS example: two of 20 from ChatGPT and seven of 20 from Gemini. This describes that reported run, not a general model ranking or a measure of all prospective customers. Search Agent OS is his implementation.
Our proposed record includes the exact question, date, language, region, session state, product name and returned answer. Give a name mention, a linked citation and an explicit recommendation separate fields. A useful audit should let someone return to the observation behind the count.
Repeat the questions before interpreting the change
Choose a modest set of genuine customer questions and repeat it on different dates. Keep branded and unbranded questions apart. Supplying the company name in the question creates a different test. Record unavailable or failed queries rather than quietly removing them from the denominator.
Improve an answerable page
Google Search Central says its Search AI features require no special SEO or special schema, and inclusion is not guaranteed. That guidance covers AI Overviews and AI Mode in Search; it does not establish rules for the Gemini app or ChatGPT.
For a proposed content trial, identify one customer question your page answers poorly. Save the original, improve its accurate and supported explanation, then observe the same question set again. Track answer mentions, website visits and actual enquiries separately. A mention can change without a corresponding change in customer demand.
Grant’s assessment: a visibility count becomes useful when its conditions are visible too. Keep the question and evidence beside the number before using it to justify more content work.
