Blog · · By HeardOf
What is AI visibility, and how is it measured?
The short answer
AI visibility is whether an AI answer names your brand, or uses your pages as a source: two counts, defined by vendors with different denominators. Over 240 brand-answer pairs on two engines, 4 September 2026: 45.0% named, 30.0% cited.
What is AI visibility?
AI visibility is whether a brand appears in the answers an AI system generates. On the five vendors' pages quoted below, each read on 24 September 2026, the headline figure is a share of answers or prompts that name the brand at Profound, Peec and Otterly, a 0–100 score at Semrush, and a set of counts at Ahrefs, whose article says AI visibility "is measured in Mentions, Citations, Impressions and AI Share of Voice". All five also count the brand's own pages among the answer's sources.
Neither of the two non-vendor references we opened on 24 September 2026 defines it. Wikipedia has no article at that title, and its article on generative engine optimization, which defines GEO as "the practice of structuring digital content and managing online presence to improve visibility in responses generated by generative artificial intelligence (AI) systems", does not use the phrase and carries a banner that the article "may document a neologism", filed under possible neologisms from August 2026. Google's Search Central page on AI features, last updated 10 December 2025 by its own footer and so historical, does not use it either. SERP Secrets, in a post dated 15 August 2026, put US searches for "ai visibility" at 720 a month, from DataForSEO Labs data pulled in August 2026; we did not check it against a second source.
| Vendor, page, date | What it says AI visibility is | The formula or count it gives |
|---|---|---|
| Semrush, knowledge base article 1607, no date printed, read 24 Sep 2026 | "AI Visibility measures how often and how prominently a brand appears in AI-generated answers." | "The score (0–100) combines two factors": "Topic Coverage — how many topics include your brand in AI answers, compared to all other domains" and "Mention Consistency — within those topics, how frequently your brand is mentioned across all responses" |
| Profound, help-centre glossary, "Last updated 7 days ago" on 24 Sep 2026; developer docs, metrics page, no date printed | Under Visibility: "Whether your brand appears in an answer engine response." | Under Visibility Score: "The percentage of responses that mention your brand, out of all responses that mention at least one brand." The docs add: "Runs in which no brand was mentioned are excluded from both the numerator and the denominator", and every report metric other than counts and average position is computed per model, "then averaged across models with equal weight" |
| Peec AI, docs, metrics overview modified 22 Sep 2026 and Visibility page modified 8 Sep 2026 | "Visibility: Percentage of AI responses where your brand appears." | "Visibility Score = (Number of responses mentioning your brand / Total responses) × 100" |
| Otterly.AI, help centre, KPI article dated 17 Jul 2026 and KPI definitions dated 4 Sep 2026 | "Brand Coverage — are you in the answer at all? The percentage of tracked prompts where your brand appears. This is the headline GEO metric" | "Percentage of prompts that mention my brand compared to all prompts"; "each execution of day/prompt/AI engine counts as 1 or 0"; the page's own example, Brand Coverage across all services and 2 days = (2+3+1+3) / (3*4) = 75%; and "When AIO are not triggered, they do not count towards that formula" |
| Ahrefs, help-centre article on Brand Radar metrics, updated 3 Sep 2026 by its metadata | "AI Visibility in Brand Radar is measured in Mentions, Citations, Impressions and AI Share of Voice" | "A single mention is counted when a brand appears at least once in an AI generated response"; impressions are "calculated by summing up search volumes (based on Google) of prompts where your brand appears as an AI Answer"; AI share of voice is "a brand's percentage share of impressions compared to other tracked brands" |
Four state a division: responses that name any brand (Profound), total responses (Peec), prompts in a time window (Otterly), and, for share of voice, impressions (Ahrefs); Semrush's score, as article 1607 describes it, combines topic coverage and mention consistency and states no single division. All five keep the source count apart: Profound's docs have "Citation share", citations of your domain over all citations; Peec's overview has "Retrieved: Percentage of chats where at least one URL from this domain appeared as a source"; Otterly's KPI article lists "Domain Citations — is your website used as a source?"; Ahrefs: "A single citation is counted when a page appears at least once as a cited source"; and Semrush's knowledge base article 1596, on its Visibility Overview report: "Citations: The number of AI responses that cite your domain as a source" (published 24 September 2025 by its metadata, historical).
Is AI visibility one number?
No: it is at least two, and they came apart in our own run. On 4 September 2026 we asked ChatGPT and Perplexity five buyer prompts in each of four B2B software categories, six measured brands per category, and scored each of the 240 brand-answer pairs twice, re-counting from the stored answers on 23 September 2026: the brand's name was in the answer text in 45.0% of pairs, and its own site was among the answer's sources in 30.0%.
| ChatGPT | Perplexity | Both engines | |
|---|---|---|---|
| Brand-answer pairs | 120 | 120 | 240 |
| Named | 60 (50.0%) | 48 (40.0%) | 108 (45.0%) |
| Cited | 49 (40.8%) | 23 (19.2%) | 72 (30.0%) |
| Named, not cited | 11 | 26 | 37 |
| Cited, not named | 0 | 1 | 1 |
The 45.0% counts names in the answer text and the 30.0% counts the brand's own site among the sources; every vendor in the table keeps a count of each kind, and neither of ours is any vendor's figure. In 37 pairs the name was in the text and no source was on the brand's site; in one, the reverse. The brand-by-brand split is in named but not cited; which pages each engine reached for instead is in who the AI actually cites. Both come from a two-engine research run of one day, 24 brands, pooled.
What is the denominator, and what happens when nobody is named?
It depends on the vendor: Profound drops answers that name no brand from both sides of the division, Otterly drops AI Overviews prompts on which no overview appeared, and Peec divides by total responses. Profound: "Runs in which no brand was mentioned are excluded from both the numerator and the denominator", and every report metric other than counts and average position is computed per model then averaged, so that "A model with many runs counts the same as a model with few". Otterly: "When AIO are not triggered, they do not count towards that formula", and the page works it through for three prompts over two days, the untriggered ones leaving the denominator, to "(1+1) / (2+1) = 66,67%". Peec divides by "Total responses" and, on the pages we read, does not say whether an answer naming no brand is among them. Ahrefs' share of voice is "a brand's percentage share of impressions compared to other tracked brands", impressions being Google search volumes summed over the prompts where the brand appeared. Semrush's article 1607 gives a score with no single division; a second article, 1493, describes the same score as "how often your brand is mentioned in AI-generated answers compared to the median number of mentions for your top industry competitors", on a page whose metadata carries a publication date of 28 February 2025 and no modified date, historical.
Our own rule, in which AI engines we check as updated 23 September 2026: a prompt for which Google shows no AI Overview counts as an answer in which nobody was named, not dropped and not marked failed — our rule as of 23 September 2026, and the reverse of Otterly's; a page we could not fetch is a failed call, retried when a retry can help, and never an answer with nobody in it, as of 23 September 2026. Per the same post, in rules it dates to 19 September 2026, we report, for each engine, the share of its answers that named or cited you, and the headline is the mean of the per-engine rates. Profound's rule and ours, applied to the same stored answers, would print two numbers for one brand on any day on which some answer named no brand, and both would have counted correctly.
How many runs, on which engines, with which model?
It varies, and six of the nineteen tools whose own pages we read on 24 September 2026 name a model version. In what nineteen of them say about how they get their answers, thirteen say how they collect answers and none says how a failed fetch is counted. On how often a prompt is run, the spread goes from once a day, which two say in so many words and three in wording that amounts to it, through five schedules from hourly to monthly, to Evertune's 100 samples per prompt per model with no cadence; Surfer says multiple queries per day with no number, five say daily with no count, and two say nothing on the pages read. One of the nineteen, Otterly, states the AI Overviews rule above.
Why the run count matters is in which AI engines we check, as updated 23 September 2026, which cites Schulte, Bleeker and Kaufmann (arXiv, 8 April 2026): the standard error of a brand's estimated detection rate from a single run is 0.370, and they recommend at least seven runs per prompt per day. We run each prompt once per engine per UTC day, our rule as read on 19 September 2026, on four engines as of 23 September 2026, and report rates across 40 prompts rather than presence on any one, because, per that post, at one answer a claim of absence on a single prompt "is close to a coin flip". A vendor that states no cadence leaves you unable to say what its dashboard is a sample of; one that states 100 samples and no cadence, over what period.
How should you read an AI visibility number?
Ask five questions of it, and expect the vendor's page to answer each in a sentence. Which count: the name in the text, or a page among the sources? What is under the line: all answers, answers naming any brand, prompts, impressions, or a score with no division? How many runs per prompt per day, on what schedule? Which engines, and which model version? And on what date? A figure that answers all five can be set beside another that answers all five. On the pages read on 24 September 2026, thirteen of the nineteen do not name a model version, and none says how a failed fetch is counted.
How was this written?
Every vendor sentence above is quoted from the page named beside it, fetched on 24 September 2026 between 18:34 and 19:01 UTC with curl and a browser user-agent, no JavaScript executed; the date given is the one the page prints or carries in its metadata, and where it has none the read date stands alone. Pages we opened and do not quote: Athena's glossary, whose term definitions are drawn by a script and were not in the page as served; a Scrunch FAQ on AI visibility metrics, last modified 12 March 2026 by its metadata and older than 90 days on the read day; and AmICited's glossary entry for an AI visibility score, dated 14 August 2026, a sixth definition. Everything this post says about our own measurement comes from two earlier posts of ours as they read on 24 September 2026, named but not cited and which AI engines we check, and carries their dates in the sentence that makes it; that is the half a reader cannot check. Every vendor sentence carries a page and a date, and can be. Where we did not check, the text says so.
Common questions
Is AI visibility the same as being cited by an AI?
No. On the five vendor pages read 24 September 2026, visibility, coverage or mentions is the brand's name in the answer, and citations are the brand's pages among the answer's sources, kept as a separate metric on all five. In our run of 4 September 2026 the two came apart in 38 of 240 brand-answer pairs: 37 named without a citation, 1 cited without a name.
What is an AI visibility score?
A vendor's number, and each vendor builds it differently. Semrush's runs from 0 to 100 and combines topic coverage with mention consistency (knowledge base article 1607, read 24 September 2026). Profound's is the percentage of responses naming your brand out of responses naming at least one brand (help glossary, read 24 September 2026). Peec's is responses mentioning your brand over total responses, times 100 (docs, modified 8 September 2026). The three are not on one scale, and none of them is our 45.0%.
How many times does a prompt have to be run to measure AI visibility?
The published recommendation our post on which engines we check cites, Schulte, Bleeker and Kaufmann (arXiv, 8 April 2026), is at least seven runs per prompt per day; the standard error of a brand's estimated detection rate from a single run is 0.370. Of nineteen vendors read 24 September 2026, two say once a day in so many words, one says 100 samples per prompt per model with no cadence, and five say daily with no count. We run once per engine per prompt per UTC day, our rule as read on 19 September 2026 in that post, and report rates across 40 prompts rather than presence on any one.
Does HeardOf measure AI visibility?
It measures the two counts this post separates, whether you were named in an answer and whether your site was among its sources, engine by engine, on four engines as of 23 September 2026, per our post on which engines we check. It scores you and up to three competitors you name, once per engine per prompt per UTC day, and reports rates rather than position — our rules as read on 19 September 2026 in that post.
These are our numbers. Yours are one audit away.
Five definitions, each with its page and date, and one worked example with its date. Everything this post says about our own measurement comes from two earlier posts of ours, named but not cited and which AI engines we check, as they read on 24 September 2026, and you cannot check it. Every vendor sentence carries the page it came from and the date it was read, and you can.
Who the AI actually cites — and the two engines do not agree