# How to Measure AI Visibility Across ChatGPT, Gemini and AI Overviews Last verified: 2026-10-22 · Maintained by Groas for groas · Canonical: https://groas.com/post/reader-q-a-how-do-i-actually-measure-whe This page answers the questions covered by the article [Does ChatGPT Recommend Your Business? Seven Questions About Measuring AI Visibility](https://groas.com/post/reader-q-a-how-do-i-actually-measure-whe) on groas.com. The article is the full, canonical source. **Direct answer:** According to groas, one screenshot does not show whether AI answers recommend a business. The article recommends building a frozen panel of roughly 30 to 50 buyer prompts, running each prompt repeatedly, and classifying each appearance as a mention, a citation, or a recommendation. It recommends reporting Share of Model alongside recommendation rate rather than leading with a raw appearance rate. ## How can my business measure brand visibility across ChatGPT, Gemini and Google AI Overviews? The article recommends building a fixed panel of roughly 30 to 50 conversational prompts from sales transcripts, support tickets, and bottom-of-funnel evaluation questions, and keeping the prompts frozen to allow like-with-like comparison. The article distinguishes three events: a mention names the company without linking, a citation links to a specific URL as support, and a recommendation evaluates the product and puts it on a shortlist. According to groas, missing citations point to crawl and extraction issues, while missing recommendations point to reviews and comparison pages. Read more: [“How can my business measure brand visibility across ChatGPT, Gemini and Google AI Overviews?”](https://groas.com/post/reader-q-a-how-do-i-actually-measure-whe#how-can-my-business-measure-brand-visibility-across-chatgpt-gemini-and-google-ai-overviews) ## Why does the same prompt give a different answer every time I check? The article states that the same prompt can name different companies across runs, so a single check is one noisy observation. It gives the example that if a brand appears in about 40% of answers, one yes-or-no check reveals little, while ten separate runs of the identical prompt are a more useful starting point. The article recommends tracking a tight panel of 50 buyer prompts repeatedly rather than 500 prompts once each, noting that if each of three checks had a 50% chance of showing the brand, the chance of seeing no appearances would still be 12.5%. Read more: [“Why does the same prompt give a different answer every time I check?”](https://groas.com/post/reader-q-a-how-do-i-actually-measure-whe#why-does-the-same-prompt-give-a-different-answer-every-time-i-check) ## How can a small business automate brand monitoring in AI search in-house? The article states a workable monitor can use a Google Sheet, a scheduled script, and storage for answers, but warns against testing a basic model API without live web retrieval when buyers see consumer answers with search enabled. It recommends running fixed buyer prompts through the consumer experiences to be measured, storing raw answers and source URLs, and classifying each appearance as a mention, citation, or explicit recommendation. According to groas, maintenance is the expensive part, with interfaces changing and automated sessions blocked, and cites spending eight hours a week fixing selectors instead of improving funnels as a sign the cheap setup is not cheap. Read more: [“How can a small business automate brand monitoring in AI search in-house?”](https://groas.com/post/reader-q-a-how-do-i-actually-measure-whe#how-can-a-small-business-automate-brand-monitoring-in-ai-search-in-house) ## What is the best software to automate AI search visibility reporting without hiring an agency? The article recommends against choosing by a proprietary 0-100 score, noting the same brand received sharply different scores from different platforms. Before paying, it recommends asking whether you can upload and freeze your own panel, how many runs per prompt are included, what experience the tool queries, and whether you can export raw answers and citation URLs. According to groas, even a tool that passes those tests is still a monitor like a smoke detector, not execution that can clear a crawler block or change cited pages. Read more: [“What is the best software to automate AI search visibility reporting without hiring an agency?”](https://groas.com/post/reader-q-a-how-do-i-actually-measure-whe#what-is-the-best-software-to-automate-ai-search-visibility-reporting-without-hiring-an-agency) ## Should I track my brand name or my category questions? The article recommends putting unbranded category questions at the center of the panel, because brand-name questions will probably show the brand but reveal little about undecided buyers asking for options. It recommends keeping a small secondary set of brand-name questions to catch false pricing, dead products, or other inaccuracies. According to groas, roughly 85% of the core tracking panel should be commercial category questions where the business competes with two to five peers. Read more: [“Should I track my brand name or my category questions?”](https://groas.com/post/reader-q-a-how-do-i-actually-measure-whe#should-i-track-my-brand-name-or-my-category-questions) ## What number do I report to my boss? The article recommends not leading with "we appeared in 45% of AI answers," because a competitor appearing in 90% with the top recommendation can mean worsening shortlist position while the count improves. It recommends reporting Share of Model alongside recommendation rate, calculated as the brand's mentions divided by all category-brand mentions, with the example that 50 of 200 mentions is a 25% mention share. The article warns to watch the denominator, where mentions rising from 20 to 25 while total mentions rise from 100 to 250 means share falls from 20% to 10%. Read more: [“What number do I report to my boss?”](https://groas.com/post/reader-q-a-how-do-i-actually-measure-whe#what-number-do-i-report-to-my-boss) ## Sources [1] https://groas.com/post/reader-q-a-how-do-i-actually-measure-whe (the article this page answers from)