Many brands do not know how often, or even whether, ChatGPT, Gemini or Google's AI features mention them. Not because it is unknowable, but because it sits outside the analytics they already have. This is how to measure AI visibility systematically and turn it into a score you can defend.
AI visibility answers the question a ranking used to answer, with one difference: instead of one position on one engine, you track a rate across several engines and repeated runs. Below is the formula, the run counts that make it trustworthy, and the bands for reading the result.
What is AI visibility?
AI visibility is how often a brand or its content appears in answers from ChatGPT, Perplexity, Google AI Overviews, AI Mode, Gemini and Copilot. It covers being named in the answer text, being cited as a source, and being described correctly.
No engine hands you a score. You build it by running real questions against each engine and recording the responses, which makes it measurable but not automatic.
How is an AI visibility score calculated?
An AI visibility score is the share of answers that mention your brand, out of every answer you collected. The denominator is prompts multiplied by engines multiplied by runs, not the number of prompts.
The formula
Visibility score = answers naming your brand ÷ (prompts × engines × runs per prompt) × 100
Twenty prompts on five engines, each run once, is 100 answers, not 20. Run each prompt three times and the denominator becomes 300. Averaging per engine instead of pooling gives a different number from the same data, so state which one you report.
Mention score versus citation score
A mention means your brand name appears in the answer text. A citation means the engine credits and links one of your pages as a source. Compute both from the same log, because the gap between them is the most useful diagnostic you have.
A worked calculation
Twenty prompts, five engines, one run each: 100 answers. Your brand is named in 22 of them and one of your pages is cited in 9.
| Metric | Count | Score |
|---|---|---|
| Total answers | 100 | Denominator |
| Answers naming the brand | 22 | Mention score 22% |
| Answers citing your domain | 9 | Citation score 9% |
| Answers naming a competitor | 61 | Competitor presence 61% |
The 13-point gap between mention and citation says the engines know the brand but reach for someone else's page when they need a source. That points at content and formatting work, not awareness.
How many runs per prompt do you need?
You need repeated runs, because the same prompt returns different answers on different runs. One run per prompt is a sample, not a measurement, and small differences between two single runs rarely mean anything.
| What you want | Runs per prompt, per engine | Prompts |
|---|---|---|
| A first look | 1 | 10 to 20 |
| An internal baseline | 3 | 25 to 50 |
| A number you will report | 5 or more, averaged | 50 to 125 |
Treat any change under a few points as noise unless it holds across two consecutive readings. The ChatGPT mention-tracking method shows the same logic on a single engine, with higher run counts for a crowded category.
How to measure AI visibility, step by step
The procedure is the same whether you run it in a spreadsheet or in software. Follow it in order.
1. Build a prompt set from real buyer questions
Start with the questions your buyers ask rather than your brand name. Sales-call notes and Search Console queries are the two best sources, and the set should cover the journey from problem awareness through to comparison.
That prompt set is your measuring instrument. If the prompts are vanity queries, the results are vanity results; comparison and recommendation prompts are where brand mentions actually happen.
2. Choose the engines that matter to you
Focus on the engines your audience uses. ChatGPT reported 800 million weekly active users in October 2025, so most programmes start there, then add Google AI Overviews and AI Mode, Perplexity, Gemini and Copilot.
Measuring an engine your audience never touches adds noise, not insight.
3. Run the prompts and record each answer
Record four things per answer: whether you appear, whether you were named or linked, which competitors appeared, and which source domains the engine used. Run in a clean, signed-out session so past conversations do not inflate your presence.
The source-domain column is where most fixes come from, so never skip it.
4. Turn the log into scores
Apply the formula above to get a mention score and a citation score, overall and per engine. Then compute share of voice: your mentions divided by all brand mentions on the same prompts, which the share of voice guide covers in detail.
5. Benchmark against competitors
A visibility figure on its own means little; 20% can be excellent or poor depending on the category. Run the identical prompt set for three to five rivals and put every brand on one scale.
The prompts where rivals appear and you do not are your content plan. This is the step where measurement turns into action.
6. Add your first-party signals
Pair prompt tracking with the free data you own. Bing Webmaster Tools' AI Performance report shows citations from Copilot and Bing's AI summaries, and your server logs show which AI crawlers reach the site.
Search Console counts AI Overview and AI Mode appearances inside overall web search traffic rather than as a separate report, per Google's documentation. In GA4, the AI Assistant default channel groups visits from ChatGPT, Gemini, Copilot and similar assistants, which shows the visits mentions turn into.
7. Freeze the set and set a cadence
Repeat the whole procedure on the same prompt set, monthly for a working baseline and every 8 to 12 weeks for a board-level trend. Change the prompts and the trend resets, which is the most common reason a team stops trusting its own number.
What counts as a good AI visibility score?
There is no absolute benchmark, because every score depends on the prompt set behind it. The only stable reading is your rate relative to the strongest competitor on the identical prompts.
Benchmark bands relative to the category leader
| Your mention score vs the leader's | Band | What it usually means |
|---|---|---|
| At or above the leader | Leading | Defend: refresh pages after each model launch |
| 50% to 99% of the leader | Competitive | Close specific prompt gaps where the leader wins |
| 20% to 49% of the leader | Challenger | Engines know you; you lack the cited pages |
| Under 20% of the leader | Invisible | Start with entity signals and third-party coverage |
These bands are a working rule for reading your own data, not an industry standard. A 20% score is "leading" if the leader sits at 18% and "challenger" if the leader sits at 70%.
Why two tools give different scores
Tools disagree for three reasons, in order of size: different prompt sets, different counting rules (mention, citation, or position-weighted) and different sampling. Ask any provider for its full prompt list before comparing its number with another's.
Why your score differs by engine
The same brand can score 30% on one engine and 5% on another, because engines retrieve differently. Google says standard SEO practices apply to its AI features, while other engines search live and cite what they find, and some answer certain questions from training data without searching at all.
Where an engine is not retrieving for your prompts, publishing more pages will not move your score there. The fix runs through third-party presence instead.
What are the common mistakes to avoid?
A few traps make AI visibility numbers misleading. Do not treat a single run as fact, and do not count only links, because a plain mention still shapes the buyer's shortlist.
Do not test your brand name instead of buyer queries, which inflates the number on questions nobody needed answered. And do not measure in a vacuum: a raw rate with no competitor comparison tells you almost nothing.
What to do once you can measure it
Measurement is only the first half. Each column in your log implies a different fix, and reading them apart stops you spending a quarter on the wrong one.
If the engine cited a competitor's page you have no equivalent of, that is a content gap, and the complete guide to answer engine optimization covers how to close it. If it cited your page but lifted a thin passage, the fix is formatting, and getting cited by ChatGPT covers the answer-first structure that survives extraction.
If it cited a third-party roundup you are absent from, no amount of on-page work helps and the job is off-site. The score is a symptom reading; the fix always comes from the layer underneath it.
How to keep the prompt set frozen between runs
Lock the prompt list in a versioned sheet or tool, and log every edit with a date so a trend break can be traced. Re-scan the same questions on live engine responses each run, because an easier question list and a rising score look identical otherwise.
Once the set outgrows a spreadsheet, the AEO and GEO tools comparison covers the software, and the GEO metrics guide defines the KPIs to report beside the score.
