AI visibility metrics: what to count, and how sure to be
Five numbers cover most of it: how often AI answers name you (mention rate), name you first (first-pick rate) and link your site (citation rate), your share of the brands they name (share of voice), and how they speak of you (sentiment). Each is a share of many answers, and answers change from one ask to the next, so each needs a range that says how sure it is.
Checked against the studies and help pages it cites on October 20, 2026.
Five metrics, and how each is counted.
Each is a share of the answers to the questions you track. What changes from one to the next is what counts as a hit, and what it's a share of.
Mention rate
Answers that name you, divided by answers that came back. A call that failed isn't an answer that left you out, so we leave failures out of every rate. Keep questions that name you apart: an answer to “is Plausible any good?” names Plausible whatever its standing.
First-pick rate
Answers that offer you as the first option, divided by answers. It stands in for rank: an average position leaves out every answer that skips you, while a first-pick rate counts them all.
Citation rate
Answers whose sources include a page on your site, divided by answers. It moves apart from the mention rate: an answer can name you without linking you, or link your page as a source without naming you as an option. The citations guide covers how ChatGPT picks its sources.
Share of voice
Your mentions, divided by mentions of you and the competitors you track. It depends on whom you track: add a rival and your share falls with no answer changed, so compare it only over one list of competitors. AI share of voice shows how GEOFast tracks it.
Sentiment
How the answers that name you speak of you: positive, neutral or negative. A language model reads each one. It's a split of the answers that name you, not of all answers.
One score, if you need one
GEOFast's AI visibility score, 0 to 100, averages a weight per answer: 1 when you're the first option, 0.8 second or third, 0.6 further down, 0.5 named outside the options, 0 not named. It sums up the first two metrics, and hides which one moved.
Why each metric needs a range.
Ask the same question twice and the answer can name other brands. So every one of these is an estimate from a sample, and needs a range that says how far off it could be.
A rate is a sample
In SparkToro's study, ChatGPT and Google's AI gave the same list of brands twice less than once in 100 tries.[1] The guide on repeat answers has the reasons. The rate over many answers holds still where one answer doesn't.
A 90% range, by the Wilson method
GEOFast gives each rate the range where the true rate likely sits, 9 times in 10, worked out by the Wilson method for proportions. NIST's handbook notes that its lower limit can't be negative, and calls an approach that can give an impossible lower limit inferior.[2]
Few answers, wide ranges
14 answers naming a brand out of 22, one engine in one run in our example below, is 64%, likely 46 to 78%. 283 out of 374 is 76%, likely 72 to 79%.
A change counts when the ranges part
GEOFast marks a move between two runs only with at least 10 answers on each side and ranges that don't overlap. Our methodology sets out the rule and every count behind it.
How many answers is enough
A preprint by IQRush, which sells AI visibility software, needed 33 to 94 cited answers per engine and topic before its ranking of cited sites stopped changing, and 3 of its 30 tests never settled within 125 questions. It is unreviewed.[3]
Share of voice's range is the tightest
It's counted over mentions, 1,460 in our example, so its range is narrow: 19%, likely 18 to 21%. Several mentions come from one answer and move together, so read that range as the least the share could be off by.
One brand's metrics, with their ranges.
We counted the five metrics for plausible.io the way the dashboard does, from the answers GEOFast recorded for it, and drew each run's mention rate with its range.
mention rate: 283 of 374 answers named Plausible, likely 72 to 79%.
first-pick rate: 157 answers offered it first, likely 38 to 46%.
citation rate: 154 answers linked a page on plausible.io, likely 37 to 45%.
share of voice against 10 tracked rivals, likely 18 to 21%. Sentiment: 263 positive, 19 neutral, 0 negative.
The runs moved from 77% to 72% and then 78%, and the ranges overlap, so none of it counts as a change. Neither does the gap between Perplexity and ChatGPT in one run.
The answers are from GEOFast's demo project for plausible.io: 22 questions that don't name the brand, asked of 6 engines (ChatGPT, Claude, Perplexity, Gemini and Google's AI Overviews and AI Mode) in US English, in 3 runs within 62 minutes on October 6, 2026: 374 answers after failed calls. Share of voice counts the 10 competitors the project tracks. Three runs that close together show how much a rate varies, not a trend. GEOFast's onboarding drafted the questions, so they aren't measured demand.
How to read the metrics without fooling yourself.
The numbers are only as good as the questions behind them and what they're compared with.
The IAB's four groups
The IAB's August 2026 framework sorts AI visibility metrics into presence (mention rate, citation rate, share of voice), prominence (placement and order), portrayal (sentiment, framing, made-up claims) and persuasion (whether visibility drives action).[4]
Directional or decision grade
The IAB calls data directional when it shows patterns, and asks for more before budget or strategy decisions: enough queries and answers, coverage of prompt types, a steady testing cadence, reproducibility and enough platforms.[4]
Sentiment on a small base
In our example, 263 of the 282 answers that named Plausible and were read for tone spoke of it positively, 93%, likely 90 to 95%. For a brand named in a handful of answers, the same split would rest on a handful.
Compare with rivals, not averages
We publish no industry averages, because we can't source them. The fair comparison is your competitors' rates on the same questions, the same engines and the same days.
Counting by hand
The five steps in the ChatGPT checking guide give you a mention rate and a first-pick rate from a spreadsheet. Keep the counts, so you can put a range on them.
Counted for you
AI visibility tracking asks your questions on a schedule and shows each metric with its count and its 90% range.
Questions about AI visibility metrics.
What are the main AI visibility metrics?
Mention rate (the share of answers that name you), first-pick rate (the share that name you as the first option), citation rate (the share that link a page on your site), share of voice (your mentions against your competitors') and sentiment (how the answers that name you speak of you). The IAB's August 2026 framework puts mention rate, citation rate and share of voice under one heading, presence.
What is a good AI mention rate?
There is no sourced benchmark to hold yours to: it depends on the questions you ask, the engines and how well known your field's brands are. Compare it with your competitors' rates on the same questions, and with your own over time, counting a change only when the two ranges don't overlap.
How many answers does an AI visibility rate need?
More than one chat can give. In our example, 14 answers naming the brand out of 22 is a rate somewhere between 46 and 78%. With 374 answers, the range narrows to 72 to 79%. A rate is worth acting on when its range is narrow enough to tell the moves you care about apart.
What is the difference between mention rate and share of voice?
Mention rate counts answers: how many of them name you. Share of voice counts mentions: yours as a share of yours and your tracked competitors'. Add a competitor to your list and your share of voice falls, though no answer changed.
How is sentiment in AI answers measured?
A language model reads each answer that names you and labels how it speaks of you: positive, neutral or negative. Sentiment is a split of the answers that name you, so it rests on fewer answers than your mention rate, and says little while you're rarely named.
Where each fact comes from.
Each numbered source is the page a fact comes from: published studies and the AI companies' own help pages, read on October 20, 2026. The answers above repeat facts marked in the guide. Help pages change, so check them before you act.
- 1SparkToro: NEW Research: AIs are highly inconsistent when recommending brands or products; marketers should take care when tracking AI visibilitysparktoro.com/blog/new-research-ais-are-highly-inconsistent-when-recommending-brands-or-products-marketers-should-take-care-when-tracking-ai-visibility/
- 2NIST/SEMATECH: e-Handbook of Statistical Methods: 7.2.4.1. Confidence intervalsitl.nist.gov/div898/handbook/prc/section2/prc241.htm
- 3Search Engine Journal: AI Visibility Rankings Aren’t Stable – New Research Shows It’s Mostly Statistical Noisesearchenginejournal.com/ai-visibility-rankings-arent-stable-new-research-shows-its-mostly-statistical-noise/
- 4IAB: IAB Releases “Measuring Visibility in the AI Era” to Help Brands, Publishers, and Agencies Navigate AI-Powered Discoveryiab.com/news/iab-releases-measuring-visibility-in-the-ai-era
Keep reading
- The free AI visibility checkerCheck your AI visibility for free: what AI crawlers can read on your site, then who ChatGPT, Perplexity and Gemini name for three of your buyers' questions.
- AI visibility tracking across six AI enginesTrack how often ChatGPT, AI Overviews, AI Mode, Gemini, Perplexity and Claude name you, put you first and cite your site, with a 90% range on every rate.
- AI visibility tools, comparedEvery tool we have checked, by starting price.
- What is GEOGEO, generative engine optimization, in plain words: how AI answers pick their sources, what research shows helps, and how to measure your AI visibility.
- AEO vs GEO vs SEOAEO vs GEO vs SEO: where each name came from, what Google and Microsoft say, the work all three share, and the four things that change with AI answers.
- Check if ChatGPT mentions your brandCheck by hand in five steps, see why one chat can't measure it (answers change, memory, location, leading questions), and where ChatGPT and Claude do help.
- Find the competitors ChatGPT recommendsA method for finding who ChatGPT recommends instead of you: which questions to ask, how many times, how to count the names and work out your share of voice.
- Does ChatGPT give everyone the same answer?No. ChatGPT writes each answer afresh and can use your memory, location, a web search and your plan's model. OpenAI's own pages, and our repeat-run numbers.
- Get ChatGPT to recommend your businessWhat gets a business named in ChatGPT's answers: what you control, whether you can submit your business, how long changes take, and how to measure it honestly.
- Get cited by ChatGPTHow ChatGPT search finds and picks the pages it links, from OpenAI's docs, what makes a page usable as a source, and which kinds of pages our data shows cited.
- Why ChatGPT doesn't show your websiteSix reasons ChatGPT leaves a site out, in the order to check them: robots.txt, firewalls, noindex, JavaScript, Bing's index and who the pages it reads name.
- Does ChatGPT use Google or Bing?What OpenAI documents about where ChatGPT search gets results (Bing, its own crawler and index), what is only reported about Google, and what it means for you.
- Rank in Google's AI OverviewsWhat Google says you need to show up in AI Overviews, how query fan-out and nosnippet work, and what studies show: 38% of cited pages rank in the top 10.
- Rank in Google AI ModeWhat sets Google's AI Mode apart from AI Overviews: query fan-out, follow-ups, a custom Gemini and Search Console's counts, and how often the two cite one page.
- Perplexity SEOHow Perplexity finds, reads and cites pages: its two crawlers, what its docs say and leave out, the Cloudflare dispute, studies, and what our runs show.
- Which AI crawlers to allowWhich AI crawlers feed ChatGPT, Claude, Perplexity and Google's answers, which only train models, and what the 1,529 of 1,605 cited sites we could read allow.
- robots.txt for AI crawlersCopy-ready robots.txt rules that keep AI search crawlers in and training crawlers out, how each crawler reads the file, and the mistakes that lock AI out.
- What is llms.txt?llms.txt is a Markdown file that tells AI agents what a site holds. Its format, a real example, who reads it today, and what studies show about its effect.
- Firewalls that block AI crawlersCloudflare, Vercel, AWS WAF and others can turn AI crawlers away even when robots.txt lets them in. Where each setting lives and how to let crawlers through.
- Do AI crawlers render JavaScript?Google's and Bing's crawlers render JavaScript. OpenAI, Anthropic and Perplexity don't say, and Vercel's 2024 study found theirs don't. Our data on cited pages.
Find out what AI says about your site.
Paste your URL and leave with a ranked list of what to fix, with steps and code.
- Free
- No account