Measure AI Visibility: 4 Metrics, Formulas, Sample Size

To measure AI visibility properly, track four numbers on every engine your buyers use: mention rate, position when named, share of voice and citation rate, plus an accuracy check. Compute them over at least 30 distinct buyer prompts, run each 3 to 5 times, and give every rate an interval. One run of one prompt is an anecdote: answers differ from run to run.

What does it mean to measure AI visibility?

Measuring AI visibility means asking AI engines your buyers' questions repeatedly and counting how often, how prominently and how accurately they name and describe your brand. The result is a rate over a defined set of prompts, engines and dates.

It is not traffic: Search Console counts traffic from Google's AI Overviews and AI Mode inside the Web search type (Google), and an unlinked mention produces no click. Track visits with GA4 referral tracking.

Which four metrics matter, and how do you calculate them?

Mention rate, position when named, share of voice and citation rate carry nearly all the signal; add an accuracy check. Count one answer per prompt, engine and run, compute per engine, and write the denominator beside every number.

Metric	Formula	Rule that keeps it honest
Mention rate	answers naming you ÷ all answers × 100	Count every run; match name, aliases and domain
Position when named	sum of your positions ÷ answers naming you	Order of first mention (1 = first); report how often you are first
Share of voice	answers naming you ÷ sum of answers naming each tracked brand (you included) × 100	Fix 3 to 6 competitors first; count a brand once per answer
Citation rate	answers citing your domain ÷ answers showing sources × 100	Count cited links, not pages merely read
Accuracy	answers naming you with no wrong fact ÷ answers naming you × 100	Check against a written fact sheet

Presence beats position. In AEOeye's audit records (small, self-selected sample; caveats below), 71.4% of 147 ranked mentions were first place, so being named at all is the hard part.

Fix the competitor set before you run anything. A set built from the answers you got leaves out the brands engines rarely name, which biases share of voice (Malthouse et al.).

Define "cited" before counting it. OpenAI's API documentation separates inline citations, which cover only the most relevant references, from a sources field listing every URL the model consulted (OpenAI).

Skip sentiment as a headline number. In AEOeye's audit records 144 of 157 mentions were positive, 12 neutral and 1 negative: invisibility, not negativity, is the common problem. Check accuracy instead; Google itself warns that AI Overviews make mistakes (Google).

Ignore the long metric lists, including our former 18-term dictionary. The rest are derivatives, sampling design or separate audits:

• Derivatives: weighted visibility, prompt coverage, engine coverage, competitor win rate and share of model. Composite scores bury their weights (ours are published).
• Sampling design (below): sample size, observation window, interval and volatility.
• Separate audits: citation correctness, completeness and source diversity; conversion rate.

For coding rules, such as what counts as a mention, see our recommendation measurement guide.

How many prompts and repeats do you need?

Use at least 30 distinct buyer prompts, run each 3 to 5 times on every engine, and put an interval on every rate. Prompts buy coverage; repeats smooth noise inside one prompt and reveal unstable prompts, your nearest wins.

A July 2026 preprint on 12,933 brand answers across three models and eight languages found that one answer carries almost no signal for telling brands apart, that repeats past the fifth barely help, and that spreading the same budget across models or languages helps several times more (Żatuchin).

Lowering temperature, the randomness setting, does not remove the noise. Five models set for deterministic output still varied by up to 15% in accuracy across ten identical runs (Atil et al.), and Anthropic's API reference says models released after Claude Opus 4.6 do not support setting it (Anthropic).

Count every run. "Named at least once in three runs" flatters you: at a true per-run rate of 30%, it happens about 66% of the time. Write prompts a buyer would type before knowing your name, each with one intent (discovery, comparison or problem), report intents separately, and lock the set.

Report rates as intervals. Use the Wilson interval, the method NIST's handbook gives for proportions (NIST), with n set to the number of distinct prompts, not answers. Repeats of one prompt are not new questions, so this errs on the safe side.

interval = (p + z²/2n ± z·√(p(1−p)/n + z²/4n²)) ÷ (1 + z²/n)
p = observed rate, n = prompts, z = 1.96 for 95%

Prompts	Named in	Rate	95% interval
30	0	0%	0% to 11.4%
30	6	20%	9.5% to 37.3%
100	20	20%	13.3% to 28.9%

Thirty prompts can tell 10% from 50% but not 20% from 30%. An interval covers noise, not a prompt set that misses your buyers' real questions.

Denominators decide what a number means. AEOeye's recommendation benchmark has 157 mentions in 892 answers (17.6%), but those answers come from 127 audits (June to September 2026) of 27 brand names, repeats included, and 509 answers come from one engine against 74 for each of three others. Four engines have under 10 mentions, and 3 in 74 carries an interval of 1.4% to 11.3%. It is self-selected, so we publish aggregate data and never compare engines from it.

How does measurement differ across ChatGPT, Perplexity, Gemini, Claude and Google AI?

Each engine decides differently whether to search, what sources it shows and how far it personalizes, so each needs written test conditions. For every answer, log the surface (app or API), date, language, location, whether search ran and account state, and never pool engines. Documentation checked 3 October 2026; apps can differ from APIs.

Engine	What changes the answer	What to fix in your test
ChatGPT	The model chooses whether to search (OpenAI)	Require search via the API, or log whether it ran
Claude	Searches for current information, including on products and organizations; the system prompt steers it (Anthropic)	Set an approximate location; log your system prompt
Gemini	The model decides if Google Search helps (Gemini API); the app can use saved info and past chats (Google help)	Turn personal context off or use a clean profile
Perplexity	Sonar Chat Completions support ended 27 September 2026; requests are being reformulated as Agent API requests (Perplexity)	Date-stamp runs; an endpoint change may break your series
Google AI	AI Overviews often do not trigger (Search Central); AI Mode can personalize from Search history (Google help)	Log whether an AI answer appeared; report mention rate among those that did

Test the tier your buyers use; AEOeye's audits run each engine at the default tier a free user would get. Split rates by searched versus not: memory-only answers move mostly with the model, search-grounded ones with the web.

How often should you re-measure, and what counts as a real change?

Re-run the full prompt set monthly, and use a weekly check of 10 to 15 prompts only as an alarm: 3 mentions in 12 prompts spans 8.9% to 53.2%. Call a change real only when the same prompts flip one way far more often than the other.

Compare prompt by prompt, because noise flips prompts both ways about equally and a lopsided count is the signal:

1. Run the same prompts every wave.
2. Mark a prompt present when you are named in at least half its runs.
3. Count gains (absent to present) and losses (present to absent).
4. Read the balance: in McNemar's exact test, 6 gains and no losses, 8 and 1, 10 and 2, or 12 and 3 clear the usual 5% bar; 5 and 2 does not.

Start a new series when the instrument changes (new default model, interface, endpoint or prompt set), log the date, and do not read the break as a brand move.

Build, buy, or audit once?

Audit once to see whether you appear at all, build a spreadsheet if you can re-run 30 or more prompts monthly, and buy a tracker for daily runs across hundreds of prompts. Insist on raw answers and written metric definitions.

AEOeye's free audit asks three unbranded buyer questions on ChatGPT and shows the full answers; the $29 report covers ten questions on five engines. Both are smoke tests for presence, not the 30-prompt design above.

One row per answer: prompt, intent, engine, surface, date, searched, named, position, linked, accuracy, answer text. At 30 prompts, 5 engines and 3 runs that is 450 rows a wave. Use an LLM judge only for fuzzy labels such as accuracy, and hand-check a sample.

Trackers win on frequency: Peec AI's documentation says it runs prompts daily (October 2026), and our tracker comparison covers the field. Check each definition, because no standard exists: Peec AI's visibility is how often your brand gets mentioned (Peec AI), while Otterly's Share of AI Voice is the percentage of citations you own versus competitors (Otterly). Different denominators make scores incomparable.