How to measure AI visibility: metrics, sample size and reporting

AI visibility is the share of AI answers in which your brand is mentioned, recommended or cited as a source. You measure it by asking 30 to 50 fixed questions in several AI systems, three times each, and scoring every answer on the same sheet. A change counts only when it is larger than the sampling error.

Key takeaways

  • In Pew Research Center's US browsing data from March 2025, users clicked a traditional search result on 8% of visits with an AI summary and on 15% without one.
  • Across ChatGPT, Perplexity, Gemini, Google AI Mode and Google AI Overviews, the top 100 cited sources of any two engines overlap by only 16% to 59%, according to BrightEdge.
  • A rate measured on 90 AI answers carries a 95% margin of about 6 percentage points at 10%; on 600 answers the margin shrinks to 2.4 points.
  • Google Search Console has shown every website its impressions in AI Overviews and AI Mode since August 31, 2026.
  • Otterly.AI's entry plan costs €29 per month for 16 prompts, fewer than a reliable measurement needs.

What is AI visibility?

AI visibility is the share of answers from ChatGPT, Gemini, Perplexity, Google AI Overviews and Google AI Mode in which your brand appears. You measure it by asking a fixed set of questions in those systems, repeating each one, and scoring every answer against the same fields.

Appearing has four levels, each with its own number: the AI mentions the brand, recommends it, links one of its pages as a source, and describes it correctly. A brand can be named in most answers and never linked, or linked and described wrongly, and a single visibility score hides both cases.

The unit of measurement is what separates this from rank tracking. A ranking is a position in a list that stays fairly stable for the same query. An AI answer is generated text that changes from one request to the next, so a sample of answers takes the place of a single rank check. For how the work on the two channels differs, see GEO vs SEO; for the discipline as a whole, see Generative Engine Optimization.

Why are clicks no longer enough?

Because a share of research now ends inside the answer. Pew Research Center analyzed the browsing data of 900 US adults in March 2025. On visits where Google showed an AI summary, users clicked a traditional search result 8% of the time, against 15% on visits without one [1]. A link inside the summary itself was clicked on 1% of visits. Around one in five Google searches that month produced an AI summary [1].

Germany shows the same direction with a different method. SISTRIX analyzed more than 100 million keywords: on average over 27% of searchers click the first result, and on keywords with an AI Overview that drops to 11% [3]. AI Overviews appeared for just over 20% of German keywords. Averaged across all keywords, the click loss was 6.6% [3]. Quote both figures together: the 27% to 11% drop applies to position 1 on keywords that show an AI Overview, and on its own it overstates the effect.

Where do AI answers come from?

AI systems answer from two places: what the model learned in training, and a live web search it runs while answering. Mentions from training carry no link and never appear in web analytics. Answers grounded in a live search cite pages, and only those citations can send visits.

For its own features, Google requires only that a page be indexed and eligible to show in Search with a snippet, with no special optimizations or schema.org markup [4]. Both features may use a query fan-out technique, issuing multiple related searches across subtopics [4]. A page can therefore be cited for a question it never ranks for in classic search.

Record two things separately in every answer: whether the brand is named, and whether one of your URLs is cited.

What do you measure?

Eight metrics cover the questions marketing, sales and leadership ask. Seven come from the same recording sheet; the eighth comes from analytics.

MetricCalculationQuestion it answers
Mention rateanswers naming the brand ÷ all answersDo we appear at all?
Recommendation rateanswers recommending the brand as a fit ÷ all answersAre we suggested?
Citation rateanswers linking at least one of our URLs ÷ all answersDo our pages serve as evidence?
Share of voiceour mentions ÷ mentions of all tracked brandsHow do we compare with competitors?
Positionaverage rank of our mention in list-style answersAre we named first or last?
Error rateanswers with a false statement about us ÷ answers that mention usIs the AI getting us wrong?
Sentimentshare of positive, neutral and qualified mentionsHow are we framed?
AI referral trafficsessions and conversions from AI sources in GA4Does visibility bring visits and leads?

A worked example. Say you ask 40 questions in five systems, three times each, and collect 600 answers. Your brand appears in 84 of them, a mention rate of 14%. Thirty answers recommend you (5%) and 18 link one of your URLs (3%). Your three main competitors collect 216 mentions between them, so your share of voice is 84 ÷ 300 = 28%. Read together, the numbers say the AI knows you, rarely recommends you and seldom uses your pages as evidence.

Run questions that contain your brand name as a separate round. Those answers almost always mention you, so the metrics that matter there are error rate and sentiment. For unbranded questions, presence comes first.

Which questions belong in the prompt set?

Start with 30 to 50 unbranded questions, phrased the way your buyers would type them into an AI system, and fix the wording once chosen. A question reworded in month three is a new question with no history.

Good questions come from three places. Search data and People Also Ask show what many people look for. Your sales team knows what prospects ask on first calls. And buyers ask AI systems full questions with context that keyword tools never capture. For this topic, Digital Twins of the audience (powered by neurotwins by neuroflash) phrased questions such as these, in German, for the German market:

  • "Welche Metriken gibt es überhaupt, um KI-Sichtbarkeit zu messen?" (Which metrics even exist to measure AI visibility?)
  • "Wie kann ich die Sichtbarkeit in KI-Antworten messen, wenn es noch keine etablierten Tools gibt?" (How can I measure visibility in AI answers when there are no established tools yet?)
  • "wie kann ich die sichtbarkeit meiner kunden in ki-suchmaschinen messen?" (How can I measure my clients' visibility in AI search engines?)

The third comes from an agency measuring for clients; the first two come from buyers who do not yet trust the tool market. A two-word keyword carries none of that context. Add comparison and buying questions with real constraints such as industry, team size or budget.

How do you run the measurement?

Ask every question in every system several times under the same conditions, and log every answer on one sheet.

  1. Pick one goal and one product. Recommendation, citation and factual accuracy each need different questions.
  2. Choose the systems: ChatGPT, Gemini, Perplexity, Google AI Overviews and Google AI Mode.
  3. Fix the conditions. New session with no history, a neutral or logged-out account, the same location, one time window per round.
  4. Run each question three times per system.
  5. Log one row per answer.
  6. Compute the metrics per system and in total, and repeat monthly with the identical set.

The recording sheet needs these columns:

FieldContent
Date, system, prompt ID, run2026-11-18, Perplexity, P07, 2
Brand mentionedyes / no
Brand recommendedyes / no
Own URL citedyes / no, which URL
Positionrank in a list, else blank
Sentimentpositive, neutral, qualified
False statementwhich, else blank
Competitors namednames
Domains citedevery source in the answer

The last column takes the most effort and pays back most, because it lists the pages AI systems use as evidence for your topic.

Is the change real or noise?

A change is real only when it is larger than the margin of error of the difference. The rule of thumb for a 95% margin on a rate is 1.96 × √(p × (1 − p) ÷ n), where p is the rate and n the number of answers [8].

AnswersMargin at a 10% rateMargin at a 50% rate
90 (15 questions × 2 systems × 3 runs)±6.2 points±10.3 points
300±3.4 points±5.7 points
600±2.4 points±4.0 points

These values are our own recomputation of that formula and assume the answers are independent.

A case published on the blog of Krauss GmbH shows how easily a trend appears from nothing. For a manufacturer of industrial doors, 15 questions were asked in 2 systems, three times each: 90 answers. The company name appeared in 9 of them, and six weeks later in 14 of 90 [8]. That reads like a rise from 10% to 16%. The difference of 5.6 points sits inside a 95% interval of about ±9.7 points for the difference. The source adds that repeated runs of the same question are not independent, so the true uncertainty is larger still [8].

This is where the size of the prompt set comes from. Forty questions in five systems with three runs give 600 answers and a margin of ±2.4 points at a 10% rate. Treat a change as real when it exceeds the margin of the difference or points the same way across several rounds. In a report, four metrics on a broad base beat twelve on a thin one.

Why track more than one AI system?

Each system draws on different sources. BrightEdge compared ChatGPT, Perplexity, Gemini, Google AI Mode and Google AI Overviews: the share of citations from authoritative sources ranges from 10% to 26% by engine [2]. The top 100 cited sources of any two engines overlap by 16% to 59%. Brand overlap is somewhat higher, at 36% to 55% [2].

Our own prompt tracking from November 2026 shows the pattern on this exact topic. We ran 15 German-language questions about measuring AI visibility in the five systems. These URLs were cited most:

URLPromptsCitationsSystems
blog.hubspot.de/marketing/aeo-kpis316Gemini, AI Mode, Perplexity
aisearch.similarweb.com/blog/geo-kpis/315Gemini, AI Mode, AI Overviews
teamlewis.com/de/magazin/ai-visibility-reporting-tools/48Gemini, AI Mode
blog.hubspot.de/marketing/tracking-ki-suche37Gemini, AI Mode, Perplexity
arxiv.org/abs/2607.1403555ChatGPT

No URL in the top ten was cited by more than three of the five systems, and the arXiv survey appeared only in ChatGPT. Tracking a single system measures that system's preferences.

What do the free data sources show?

Google Analytics 4, Google Search Console and Bing Webmaster Tools show clicks and the Google and Microsoft side of AI visibility. What ChatGPT, Perplexity or Gemini say about you without a click stays out of view.

Google Analytics 4. Create a custom channel group for AI assistants under Admin, Data display, Channel groups [7]. Standard properties can have two custom channel groups. Order matters, because each session goes to the first channel whose definition it matches [7], so place the AI channel above Referral. An example source pattern:

chatgpt\.com|chat\.openai\.com|perplexity\.ai|gemini\.google\.com|copilot\.microsoft\.com|claude\.ai

Google Search Console. Google launched Search generative AI performance reports on June 3, 2026, and rolled them out to all websites worldwide on August 31, 2026 [5]. They cover AI Overviews, AI Mode and generative AI features in Discover. An impression counts each time a URL from your site appeared in one of those features, with breakdowns by page, country, device and date [5]. AI Overviews and AI Mode also remain inside the standard Performance report under the Web search type [4].

Bing Webmaster Tools. Since February 2026, a public preview shows how your content appears in Microsoft Copilot, AI summaries in Bing and select partner integrations [6]. It reports total citations, average cited pages, grounding queries, page-level citation activity and trends over time [6]. The grounding queries are worth feeding back into your prompt set.

Together, these sources tell you whether your pages are cited and bring visits. Only your own runs tell you whether an AI recommends you and what it says.

What does measurement cost?

By hand it costs time; with a tool it starts at €29 per month, and the prompt allowance decides whether the numbers hold up.

On November 18, 2026, Otterly.AI listed Lite at €29 per month with 16 prompts, Standard at €189 with 120 prompts and Premium at €489 with 403 prompts [9]. The entry plan covers fewer than half of the 40 questions a ±2.4-point measurement needs, so use it as a trial and avoid month-on-month comparisons from it. Semrush listed its SEO plan at €134 per month on monthly billing, or €112.52 per month billed annually, with AI search features bundled in [10]. The pricing page showed no standalone price for AI visibility.

For manual work, an estimate of our own: 600 answers at an assumed one minute each for running and logging come to about ten hours per round, before analysis. Before buying a tool, check four things: prompts included, systems queried, runs per prompt, and whether you can export the raw answers. Without raw answers you cannot check errors or recompute the margin. Prices change; treat these as of the date given.

How do you report AI visibility to a client or board?

One page per month works when it shows four metrics with their margins, the competition, real answers and next steps.

  • Mention rate, recommendation rate, citation rate and share of voice against last month, each with its interval. A change inside the interval is reported as unchanged.
  • Which competitors are named instead of you, and for which questions.
  • Three real answers: a strong one, a weak one and, if any, one with a false statement. A board reads an actual answer faster than a rate.
  • The domains the systems cite for your topic, as targets for content and PR.
  • The changes planned for next month and the metric each one should move.

Report AI referral traffic and its conversions as the business result, and the other metrics as leading indicators.

How GEO Agents runs this

In the GEO Agents managed service, agents do the recurring work described here. They build the prompt set from search data and from the questions Digital Twins of the audience ask, run it repeatedly in ChatGPT, Gemini, Perplexity, Google AI Overviews and Google AI Mode, and compute these metrics. A person releases every content change that follows before it goes live. The full loop is on How it works.

What can measurement not tell you?

Every number holds only for your prompt set, your period and your method. Answers cannot be reproduced exactly, a model update can shift values overnight, and figures from different tools are not comparable because each uses its own prompts, systems and repetitions. Compare your own trend with one method.

Cause and effect is the harder limit. A survey by Olivier Martinez, a preprint from July 2026 that has not been peer reviewed, concludes that "no reviewed technique shows a stable, longitudinal, cross-platform causal effect on organic discoverability" [11]. The author describes GEO as "a stochastic, partially observable pipeline" and proposes measuring discoverability, citation, absorption and economic outcomes separately [11]. That matches the method above. Measure each level on its own, report the margin with every number, and treat any link between an action and a metric as a hypothesis until several rounds support it.

Questions and answers

What metrics should I use to measure AI visibility?

Track mention rate, recommendation rate, citation rate, share of voice, position in the answer, error rate, sentiment and AI referral traffic from Google Analytics 4. For a monthly report the first four are usually enough, each shown with its margin of error and next to two or three competitors.

How many prompts do I need to measure AI visibility reliably?

Work back from the number of answers. Forty prompts in five AI systems with three runs each give 600 answers, which puts the margin of error on a 10% rate at about 2.4 points. Fifteen prompts in two systems leave a margin of about 6 points, so small changes stay inside the noise.

Can I measure AI visibility for free?

Partly. Google Analytics 4 shows visits from AI assistants, Google Search Console shows impressions in AI Overviews and AI Mode, and Bing Webmaster Tools shows citations in Copilot. Mentions without a click in ChatGPT, Perplexity or Gemini only show up when you run the questions yourself.

How do I know if my AI visibility actually improved?

Compare the difference between two measurements with its margin of error. A brand that moves from 9 to 14 mentions out of 90 answers gained 5.6 points, while the 95% interval of that difference is about plus or minus 9.7 points. That is noise. More answers or a consistent trend across several rounds are needed.

Sources

  1. 1
    Google users are less likely to click on links when an AI summary appears in the resultsPew Research Center, accessed pewresearch.org
  2. 2
    Why AI Engines Cite Different Sources but Recommend the Same BrandsBrightEdge, accessed brightedge.com
  3. 3
  4. 4
    AI features and your websiteGoogle Search Central, accessed developers.google.com
  5. 5
    Search Generative AI performance reportsGoogle Search Central Blog, accessed developers.google.com
  6. 6
    Introducing AI Performance in Bing Webmaster Tools Public PreviewBing Webmaster Blog, accessed blogs.bing.com
  7. 7
    Custom channel groupsGoogle Analytics Help, accessed support.google.com
  8. 8
    Messung der KI-Sichtbarkeit: Methoden und Praxis für präzise AnalyseKrauss GmbH, accessed krauss-gmbh.com
  9. 9
    Otterly.AI PricingOtterly.AI, accessed otterly.ai
  10. 10
    Semrush PricesSemrush, accessed semrush.com
  11. 11

Powered by neurotwins, the Digital Twins by neuroflash

Search data by KeywordCluster

Buyer questions in this article come from Digital Twins of the audience and from search data.

Want your brand in these answers?

Book an intro call. We look at the prompts that matter in your market.

Book an intro call (opens in a new tab)