AI visibility metrics, explained: how to measure your brand in AI search
The six numbers that describe how AI assistants see a brand: mention rate, share of voice, position, citation share, sentiment and accuracy. What each one means, how to calculate it and how not to misread it.
Short answer: AI visibility is measured by asking AI assistants a fixed set of questions many times and counting. Six metrics cover it: mention rate (how often you are named), share of voice (your share of all brand mentions), position (how early you are named), citation share (how often your pages are cited as sources), sentiment (how positively AI describes you) and accuracy (whether what it says about you is true).
The six metrics at a glance
| Metric | Question it answers | How to calculate |
|---|---|---|
| Mention rate | Does AI recommend us? | Answers that name you ÷ all answers, per question and engine |
| Share of voice | How big are we next to competitors? | Your mentions ÷ all brand mentions across the question set |
| Position | Are we the first choice or an afterthought? | Share of answers where you are named first, or your average rank |
| Citation share | Does AI use our pages as evidence? | Answers citing your domain ÷ answers that cite any source |
| Sentiment | Does AI speak well of us? | Score from 0 (negative) to 100 (positive) on questions about you |
| Accuracy | Is what AI says about us true? | Correct factual claims ÷ all factual claims about you |
Mention rate
The core metric: in what share of answers does AI name you? Measure it per question and per engine, because they differ a lot. A brand can be named in most ChatGPT answers and almost none from Gemini for the same question, usually because one engine searched the web and the other answered from memory. In our beauty study, ChatGPT searched the web before 87% of its answers and Gemini before 5% (Orbie, 2026).
Share of voice
Your mentions as a share of all brand mentions in the answers. It is useful for comparing yourself with competitors, but it depends on how many brands each engine lists. In our scan of questions about AI visibility tools, Gemini named 3.9 companies per answer on average and ChatGPT 1.9 (Orbie, 2026). The same brand gets a smaller share of a longer list, so compare share of voice within one engine, not across engines.
Position
Being named first matters more than being fifth in a list of eight. Track the share of answers where you are named first, or your average position when named. Ignore position in answers that don't rank anything.
Citation share
When an engine searches the web, it cites the pages it used. Citation share is how often those citations include your own site. It tells you whether your pages are seen as evidence, and the cited pages of competitors and third parties show where your brand is missing. See what makes content get cited by AI.
Sentiment and accuracy
These come from questions that name you: 'Is [brand] worth it?', '[brand] reviews', '[brand] vs [competitor]'. Sentiment scores the tone across repeated answers, and the themes behind it (the strengths and concerns AI keeps repeating) are more useful than the score alone. Accuracy checks each factual claim (prices, products, locations, policies) against your own site. A wrong fact repeated in many answers is often the most urgent finding. More on AI brand sentiment.
Five mistakes that make the numbers wrong
- Measuring one answer. Answers vary from run to run; count over repeated answers and report a range. 3 of 10 answers means a true rate somewhere between about 11% and 60% (95% interval).
- Mixing engines into one number. An average across engines hides the engine where you are missing.
- Choosing questions that flatter you. Questions that contain your name or your exact product terms inflate every metric. Use the customer's words.
- Counting answers with no brands as losses. Some answers give general advice and name nobody. Report them separately.
- Comparing over time with different questions. A before/after comparison is only fair with the same questions, engines, settings and number of runs.
How often to measure
Answers that rely on live search can change within weeks of new pages being indexed; answers from a model's memory change mainly with new model versions. Measuring every day with the same question set shows real movement: one day's answers vary, but the trend over a few weeks doesn't. That is how monitoring works.