AI visibility is how often, in what position and on the strength of which sources your brand is mentioned in the answers of AI assistants such as ChatGPT, Gemini, Perplexity and Google AI Mode. There is no equivalent of the rank-tracking report you know from SEO, so you have to build the measurement yourself. The method below is the measurement layer Roicool uses in its GEO work, and any team can run it.
Why a classic SEO report isn't enough
A search results page has a ranking for every query, and tools track it daily. An AI answer has no fixed list. It is generated fresh each time, brands are mentioned inside the text, and the cited links change from one answer to the next. A page can rank well and still not be mentioned in the answer; a page outside the top 100 can still be cited. Ahrefs found that only 38% of Google AI Overview citations come from pages ranking in the top 10 for the same query (Ahrefs, 2026).
That gap is why "we rank on page one" doesn't answer the question your leadership is starting to ask: "Does ChatGPT recommend us?" Answering it takes a different kind of record.
We run this measurement with Ossvy, the software we built; the method below works with any tool, or with a spreadsheet.
Step 1: Build the question set
Everything rests on a set of questions that stays the same every month. It is made of four types:
| Question type | Example | What it shows |
|---|---|---|
| Brand | "What is Roicool and what does it do?" | Whether the model knows your brand correctly |
| Category | "Agencies offering GEO services in Turkey" | Whether you're mentioned in category searches |
| Comparison | "HubSpot or Zoho: which is better for B2B?" | Which sources lead at the moment of decision |
| Problem | "Why are Google Ads conversions missing?" | Whether you're cited in your area of expertise |
For one service area, 20–40 questions is a good start. Write them in the words your customers actually use. Sales calls, support tickets and Search Console queries are the best raw material: long, question-shaped queries in Search Console are often close to how people phrase a prompt.
A few practical rules help the set last:
- Keep the brand questions few. Two or three are enough to check that the model describes you correctly. The real demand sits in category and problem questions, where nobody types your name.
- Write category questions the way a buyer would. "Which agencies in Istanbul do B2B lead generation?" is closer to a real prompt than "lead generation agency".
- Include your competitors' territory. Comparison questions show which brands the model already puts side by side, and which sources it leans on to do so.
- Freeze a core set. New questions will come up. Track them in a separate group so the core set stays comparable month to month.
Step 2: Choose the platforms and how often to repeat
Run the set on at least four platforms: ChatGPT, Gemini, Perplexity and Google AI Mode. Because answers are not deterministic, it is more reliable to run each question more than once and record the result as a rate. The history of a signed-in account can colour the answers, so use a clean session or a tracking tool where you can, and keep location and language fixed.
Write down the conditions of each run (date, platform, language, location, signed in or not) so next month's run can repeat them. A change in conditions looks exactly like a change in visibility in the data, and you won't be able to tell them apart afterwards.
Step 3: Keep the log
Each answer is one row. The columns we recommend:
| Column | What goes in it |
|---|---|
| Date | The day the measurement was taken |
| Question and type | The question from the set and which of the four types it is |
| Platform | ChatGPT, Gemini, Perplexity, AI Mode |
| Mentioned | Does the brand appear in the answer (yes / no) |
| Position | If several brands are listed, where you appear |
| Cited source | The links the answer relies on |
| Description accurate | Is the brand described correctly |
| Competitors | Other brands mentioned in the same answer |
Four metrics come out of this log: mention rate (the share of answers that mention the brand), average position, citation share (how often your own site, or pages about you, are cited as a source) and competitor share (how often competitors are mentioned on the same questions).
To make that concrete: if you run 30 questions three times each on four platforms, you have 360 rows a month. Mention rate is simply the rows with "yes" divided by 360, and you can break it down by question type and platform. A brand that shows up well on brand questions but rarely on category questions has a different problem from one that is mentioned often but described wrongly.
Step 4: Separate the clicks in GA4
The question set shows visibility; GA4 shows the visits that visibility brings. Group sessions whose source is chatgpt.com, perplexity.ai, gemini.google.com, copilot.microsoft.com or claude.ai into a segment or a custom channel group. Report this segment's conversions separately; visitors from AI tools often arrive near the end of their research.
There's a limit: GA4 only sees clicks. Only the question set measures a brand that is mentioned in an answer but never clicked.
Step 5: Add Search Console
Google says pages that appear in AI features such as AI Overviews and AI Mode are counted within overall web search data in the Search Console Performance report. With no separate filter, isolating that data is limited; even so, rising impressions on long, question-shaped queries are a useful signal.
Common measurement mistakes
Most errors come from the method, not the platform.
- Changing the set every month. Adding new questions is natural, but the core set has to stay fixed for the comparison to mean anything. Track new questions in a separate group.
- Asking only brand questions. A model can know your brand and still never mention it on a category question. Real demand lives in category and problem questions.
- Looking at a single answer. Missing from one answer tells you nothing; how many of ten answers mention you tells you a lot.
- Not logging the cited source. If you don't know which pages are cited, you don't know what to fix.
- Skipping description accuracy. Being mentioned isn't enough; your brand needs to come up with the right service, the right city and the right audience.
- Measuring from a personal account. An account with history can personalise answers; measure from a clean session.
How to read the results
One month's result is a baseline; it doesn't support a decision on its own. In months two and three, ask: on which question type did the mention rate move? Are there new directories or lists among the cited sources? Has a wrong description been corrected? Where did competitor share grow? Platforms change how they pick sources often, so review the method every quarter.
A simple monthly report needs four things: the mention rate by question type and platform against last month, the top cited sources for your category questions, any wrong descriptions found, and the AI referral segment's visits and conversions from GA4. That's enough for a team to decide what to fix next.
If you want to set up this measurement for the first time, Roicool's AI visibility analysis establishes the baseline across four platforms and ends with a roadmap. The work that follows the measurement is described on our GEO agency page, and the steps that move the numbers are collected in our GEO checklist.
Sources
- Ahrefs, 38% of AI Overview citations pull from the top 10 (2 March 2026): ahrefs.com/blog/ai-overview-citations-top-10 (accessed October 2026)
- Google Search Central, AI features and your website: developers.google.com/search/docs/appearance/ai-features (accessed October 2026)
- Aggarwal et al., GEO: Generative Engine Optimization (KDD 2024): arxiv.org/abs/2311.09735 (accessed October 2026)
Frequently asked questions
How many questions do I need for a reliable measurement?
For one service area, 20–40 questions is a good start: a few each of brand, category, comparison and problem questions. What matters is not the size of the set but that it stays the same every month; you can't compare two months measured with different sets.
If the same question gets different answers, is the measurement reliable?
AI answers are not deterministic; the same question can get a different answer on the same day. That's why you look at rates, not single answers: run each question several times and record the share of answers that mention your brand.
How do I see AI traffic in GA4?
Group sessions whose source is chatgpt.com, perplexity.ai, gemini.google.com, copilot.microsoft.com or claude.ai into a segment or a custom channel group. It only shows visits that came from a click; it can't show visibility in answers that mention you without a click.
Is there Google AI Mode and AI Overviews data in Search Console?
Google says pages appearing in AI features such as AI Overviews and AI Mode are counted within overall web search traffic in the Search Console Performance report. There is no separate filter, so isolating that data on its own is limited.
Should I measure with a tool or by hand?
Either works. Measuring by hand is enough to start with a small set; as the set and the number of platforms grow, a tracking tool saves time. Whichever you choose, keep the log structure the same: question, platform, mention, position, cited source, date.
Sources
Zehra AdaşManaging Partner
Zehra Adaş is Roicool's co-founder and a Managing Partner, and leads strategic marketing, specializing in brand positioning, digital communication and creative strategy.
