Guides
What is AI visibility? How it is measured and what it misses
AI visibility is the share of collected AI answers naming your brand. It is a sample: about 300 answers per period land it at roughly ±5 percentage points.

By Mathias C.SEO consultant, and more recently GEO, at MentionBox
AI visibility is the share of AI answers that name your brand, inside a set of answers somebody collected. A measurement tool sends a list of prompts to ChatGPT, Perplexity, Gemini, Google's AI surfaces or a model API, stores what comes back, counts the answers naming you and divides by the answers collected. The engines publish none of this: the responses documented by OpenAI, Anthropic and Perplexity carry the answer text, the searches behind it and the URLs consulted, with no visibility field anywhere. Because the figure is counted over a sample, its precision depends on the size of that sample: cloro's study of August 14, 2026 puts about 300 answers per period at roughly ±5 percentage points, and about 294 answers on each side before a claimed move from 20% to 30% is a result.
This page is for whoever has to present one of these percentages, and for whoever writes the code that produces it. It covers the three choices that decide the number, the facts a stored answer really settles, the gap between a citation and a mention, the questions the metric cannot answer at any sample size, and the checks to run before a visibility figure goes into a report.
The three choices that produce the number
Everything after these three decisions is arithmetic.
- The prompt list. You choose it, and no engine publishes how often anyone asks those questions. Vendors that weight prompts say they model the weight: Ahrefs builds AI impressions by "summing up search volumes (based on Google) of prompts where your brand appears as an AI Answer", and DataForSEO's AI Keyword Data API returns volumes "reflecting their estimated usage in AI tools". Add ten prompts to the list and every percentage computed over it moves, with nothing having changed in any answer.
- The engines and the locale. A brand named in most Perplexity answers and absent from Google AI Overviews has two visibility figures, and an all-engine average hides both. The country and language of the run change the answer set again.
- The number of runs. One run is a draw from a distribution. SparkToro's research, in which 600 volunteers ran 12 recommendation prompts through ChatGPT, Claude and Google's AI for 2,961 combined runs, reports under a 1 in 100 chance that two responses carry the same list of brands, and about 1 in 1,000 before two lists arrive in the same order.
What one collected answer settles, and what it does not
| Question about one answer | Settled by the stored answer? | What supplies it |
|---|---|---|
| Was your brand named in the text? | Yes | the answer text, matched against your brand and its aliases |
| Where in the text was it named, and with which competitors? | Yes | the order of brand names inside the answer |
| Was one of your URLs cited as a source? | Yes | the citation and source fields of the engine response |
| In what tone was the brand described? | Yes, as a vendor score | sentiment computed by the measurement tool |
| Which searches did the engine run first? | On some engines | the fan-out queries, when the engine exposes them |
| How many people asked that question? | No | nothing publishes it |
| Did anyone click through to your site? | No | your analytics, or a first-party report |
| Did a real user see this exact answer? | No | the answer was generated for the tool, in its own session |
Fields from the engine API references linked in this article, read on November 14, 2026.
The first four rows are what a visibility report is built on, and they are genuine observations. API for GEO is a readable example of the computed layer on top of them: its AI Visibility API returns visibility as the "percentage of in-scope responses mentioning the entity", shareOfVoice as "entity mentions ÷ all entity mentions in scope", avgPosition as the "mean rank of the first mention inside the answer" and a sentiment score per entity, at $13 per 1,000 answers down to $7 at its largest tier, checked on October 23, 2026. MentionBox, which publishes this site, has a commercial link with API for GEO. The formulas behind those names differ from one vendor to the next, which our guide on share of voice, visibility and position works through field by field.
A citation and a mention are two different counts
The same answer can link to your page without saying your name, or say your name without linking to anything. Semrush's study of June 9, 2026, built on 3,981 domain appearances across 115 prompts in 14 countries on ChatGPT, Gemini, Google AI Overviews and Google AI Mode, puts numbers on the split: 74.9% of brand appearances carried a citation, 38.3% carried a brand mention, and 62% of citations arrived with no mention of the brand in the answer text.
The split moves by engine in the same study: Gemini named brands in 83.7% of appearances while citing in 21.4% of them, and ChatGPT cited in 87% while naming the brand in 20.7%. A single "AI visibility" score that folds mentions and citations together therefore reads differently depending on which engines dominate the sample, and two brands with the same score can have opposite profiles. Keep the two counts in separate columns; our guide on tracking AI citations of your pages covers the citation side on its own.
What AI visibility does not measure
Exposure. A visibility percentage treats every prompt in your list as equally important, because nothing tells you how often each one is asked. The figure answers "when this kind of question is asked, how often does this brand come up", which is a survey result rather than an impression count.
What any given person saw. Collection runs in clean sessions, and consumer surfaces are personalized. Google announced Personal Intelligence in AI Mode on January 22, 2026, an opt-in Labs feature for AI Pro and Ultra subscribers in English in the US that connects Gmail and Google Photos to Search. On the API side the asymmetry is explicit: OpenAI documents that "each text generation request is independent and stateless", so a measured answer carries no history, while a logged-in user's does.
Clicks and revenue. The free first-party reports stop at exposure to a link. Search Console's generative AI performance report, rolled out to all websites worldwide on August 31, 2026, shows impressions for AI Overviews and AI Mode with no clicks, no queries and no position. Bing Webmaster Tools reports citations and citation share per grounding query, also without clicks. On the behaviour behind those links, Pew Research measured 68,879 Google searches by 900 US adults in March 2025: users clicked a traditional result link in 8% of visits where an AI summary appeared, against 15% where none did, and clicked a link inside the summary in 1% of visits.
Whether the engine searched at all. Anthropic's web search documentation states that Claude "answers directly without searching when the request draws on stable knowledge", and that triggering "is steerable through your system prompt". An answer written from model memory and an answer assembled from fresh search results look alike in a visibility count and come from two different mechanisms, only one of which your pages can reach. Query fan-out is the step that decides the second case.
Why you were absent. A missing brand and a missing answer are the same empty cell. Google states that AI Overviews "are only shown when our systems determine that it is additive to classic Search, and as such, often don't trigger", so a zero can mean the surface never appeared.
Causality. No AI visibility dataset contains a person. It cannot attribute a signup to a mention, and in our view any report that implies it should be read as a model with its assumptions written down.
Why two tools report two numbers for the same brand
Part of the gap comes from formulas, and part from the surface that was sampled. The two main collection routes return different objects: the consumer interface shows what users see, with its interface rules and its logged-out defaults, while an official model API runs the model you asked for, with no product layer around it.
| Collection method | APIs on this site | What the answers represent |
|---|---|---|
| Live consumer interface only | 25 of 55 | the product your buyers use, scraped session by session |
| Interface and an official model API | 12 of 55 | both, depending on the engine requested |
| Official model API only | 2 of 55 | the model's output under your own parameters |
| Method not published | 16 of 55 | unknown, which makes the figure hard to defend |
Counts from the 55 AI visibility APIs compared on this site, checked on October 23, 2026.
Two tools sampling two surfaces, with two prompt lists, two run counts and two denominators, produce two valid numbers for the same brand in the same month. Our comparison of rank tracking and GEO tracking covers how far the data model drifts from the SEO table most teams still report beside it.
Five checks before a visibility figure goes in a report
- State the denominator in one line. Answers naming your brand, divided by answers collected, with both counts stored rather than the percentage alone.
- Name the engines and their weights. An average across four engines with 400, 300, 200 and 50 answers is not a plain mean of four percentages.
- Write down the run count. Compare it with the gap you claim to detect, using cloro's figures as the yardstick.
- Freeze the prompt list for the length of the comparison. Changing it mid-quarter moves every percentage on its own.
- Report mentions and citations separately. The Semrush split shows how far the two can diverge inside one engine.
Our view
In our view AI visibility is a useful measurement as long as it is read as a survey: it says how often a brand comes up when a category is discussed, on the sample you defined, with a margin of error that comes from the run count. It is not a ranking, not an audience figure and not a traffic forecast, and presenting it next to clicks in the same chart invites a comparison the data does not support.
Frequently asked questions
What is AI visibility?
AI visibility is the share of AI answers, within a set someone collected, that name your brand. A tool sends a list of prompts to engines such as ChatGPT, Perplexity, Gemini or Google's AI surfaces, stores the answers, counts the ones naming you and divides by the answers collected. The engine returns none of this: the published API responses from OpenAI, Anthropic and Perplexity carry the answer text and its sources, with no visibility field and no rank.
How is AI visibility measured?
By repetition over a fixed prompt list. Three decisions produce the number: which prompts you track, which engines and countries you run them on, and how many times per period you run each one. One run settles nothing, since SparkToro's research across 2,961 runs of 12 prompts found under a 1 in 100 chance of the same brand list appearing twice, and cloro puts about 300 answers per period at roughly ±5 percentage points.
Does AI visibility measure traffic from AI answers?
No. A visibility percentage counts appearances in answers, and an appearance can arrive with no link at all. Search Console's generative AI performance report, live for all sites worldwide since August 31, 2026, shows impressions for AI Overviews and AI Mode with no clicks, no queries and no position, while Bing Webmaster Tools reports citations and citation share per grounding query, also without clicks. Referral traffic from AI answers is a separate measurement taken in your analytics.
Can an AI visibility tool see what real users are asking?
No engine publishes prompt volumes, so the prompt list is a sample you chose rather than a map of demand. Vendors that attach weight to prompts model it: Ahrefs builds AI impressions by summing the Google search volumes of prompts where a brand appears, and DataForSEO's AI Keyword Data API returns volumes reflecting their estimated usage in AI tools. Both say so in their own documentation.
Why do two AI visibility tools report different scores for the same brand?
Because they count different populations. Denominators differ by vendor, prompt lists and engines differ, and so does the surface sampled: of the 55 AI visibility APIs compared on this site, 25 collect from the live consumer interface, 12 from both an interface and an official model API, 2 from a model API alone and 16 do not publish the method, checked on October 23, 2026. Two numbers built on two surfaces are two measurements, not a disagreement.
Is AI visibility the same as a ranking position?
No, because an AI answer has no ordered result list to read a position from. What vendors call position is the rank of your brand's first mention among the brands named inside the answer, averaged over the answers where you appear, so it improves when your appearances fall. Visibility is binary per answer: your brand is named or it is not, and the reportable figure is the rate over many answers.