Academy · Updated 2026-08-23 · 7 min read
How to measure brand visibility in AI answers
The board asked how the brand shows up in AI answers and nobody has a number. Here is the framework: seven metrics, how each is computed honestly, and the traps that make most AI-visibility numbers wrong.
"How is our brand represented in AI answers?" is now a board question, and it has a measurable answer — but only if the measurement respects what answer engines are: probabilistic systems that give different answers to the same question, grounded sometimes and not others. This is the framework AEOSearch reports are built on; it works with any tooling that stores its evidence.
The seven metrics
1 · Mention rate. Of all stored answers to your question set, the share that name your brand at all. Reported per engine — the engines differ, and the differences are informative.
2 · Unprompted mention rate. The same, restricted to questions that name no brand — phrased the way a cold buyer asks. This is the headline number, and it is the one most reporting quietly inflates by blending in branded questions. Ours are scored separately by construction.
3 · Share of AI voice. Your brand's share of all tracked-brand mentions across the answer set — the zero-sum scoreboard against named competitors, which is what a board actually wants to see moving.
4 · Average rank when mentioned. Being named first and named fourth are different outcomes; this catches the difference the mention rate hides.
5 · Recommendation rate. Mentioned is not endorsed. The extraction marks which mentions are actual recommendations, and the gap between the two numbers is real — in our own baseline, a quarter of our mentions were recommendations.
6 · Citation share. How often the engines cite your domain as a source. This is the evidence-pool metric: it can be zero while mention rate is positive, and that combination tells you the model knows of you but never reads you.
7 · Recognition boundary. Mention rate laddered by question specificity — from questions that name you down to generic category questions. The deepest rung where you still get named is where recognition ends; the monthly goal is pushing it one rung down.
The three traps that invalidate the number
Single-run sampling. The same engine answers the same question differently run to run; a number from one pass is noise. Sample every question multiple times (we run three per engine) and report from the full set of stored answers.
Ignoring grounding. An answer built from live web search reflects today's evidence pool; an ungrounded answer reflects training data. They move on different timescales and respond to different fixes, so they must be labeled — the grounding rule is the longer argument.
Unstored evidence. If the answers behind a metric aren't stored and readable verbatim, the metric cannot be audited — not by a client, not by the board, not by you in three months. Every AEOSearch figure traces to a stored answer; hold any measurement, ours included, to that bar.
What to tell the board
Report four things on a monthly cadence: unprompted mention rate and share of AI voice (the scoreboard), citation share (the evidence pool), and — for a brand starting from zero — time to first unprompted mention as the milestone metric. Attach one stored answer per lost question; verbatim answers do more for a board's understanding than any chart.
We publish our own numbers because we ask you to trust ours. On 2026-08-23 we ran the full AEOSearch audit on AEOSearch itself — 300 stored, grounded answers across ChatGPT, Claude, Gemini, Grok and Perplexity. Our unprompted mention rate was 0%. Our citation share was 0%. That is the honest starting line for a young brand, and every method on this page is what we are running on ourselves, in public, to move it.
The fastest way to have this framework populated with your brand's numbers is the free scorecard — 8 questions, real engines, stored answers — with the $199 full audit behind it for the complete question set and the fix plan.
Questions, answered straight
How do I measure how often ChatGPT mentions my brand?
Build a fixed question set your buyers actually ask, put it to the engine repeatedly with live search on — single runs are noise — and store every answer. Mention rate is the share of stored answers naming your brand; keep questions that name no brand separated, because that unprompted rate is the honest headline number.
What is share of AI voice?
Your brand's share of all tracked-brand mentions across a stored answer set — a zero-sum scoreboard against named competitors. Unlike raw mention rate it moves when a rival gains even if you hold steady, which is why it is the metric closest to what boards mean when they ask about being represented.
What's a good AI visibility score?
There is no cross-industry benchmark worth trusting, and vendors quoting one are inventing it. What is meaningful: your trend against your own stored baseline, and your share of voice against named competitors on your actual buyer questions. For young brands the milestone metric is time to first unprompted mention.
Why is my brand mentioned but never cited?
Mentions come from the model's learned associations; citations come from the live evidence pool it reads while answering. Named-but-never-cited means the model knows of you but your pages aren't among its sources — check crawler access and whether you have citable, extractable pages for the questions in play.
— The AEOSearch team