The GEO metrics worth tracking (and the misleading ones)
Mention rate, citations, share of voice, visits and enquiries from AI: how to read them with their margin of error, and which numbers to ignore.
By Sébastien Monnier · · 5 min read · Lire en français
How often are you recommended? With a link or without? By which assistant? And does it bring you customers? Here are the metrics that answer those questions, the ones that mislead, and how to read them without fooling yourself.
The useful metrics
| Metric | What it measures | How to read it |
|---|---|---|
| Mention rate | Share of answers that name your business | Always with its margin of error (95% interval) |
| Citation rate | Share of answers with a link to your site | This is what can produce a click |
| Weighted visibility | Mention rate, weighted by the demand behind each question | A question people ask a lot counts more |
| Share of voice | Your share among all the businesses cited | Compare it with your direct competitors |
| Average position | Where your name appears in the answer | Cited first = more visible |
| Visits from AI assistants | Measured clicks to your site from ChatGPT, Perplexity… | A floor: some clicks cannot be identified |
| Enquiries from AI assistants | Calls, emails, forms after a visit from an AI assistant | The most direct link with revenue |
Table compiled on 26 September 2026.
The margin of error is not a detail
A 50% rate measured on 3 answers may in reality lie anywhere between 9% and 91%. On 300 answers, between 44% and 56%. That is why each question is asked several times, on each assistant, in every measurement.
A simple rule: an increase is real only if the new measurement’s interval lies entirely above the reference measurement’s interval. Otherwise it is “within the margin of error”.
The misleading metrics
| Misleading metric | Why |
|---|---|
| A “we’re cited!” screenshot | One isolated answer, impossible to reproduce |
| An unexplained “AI visibility score” | You don’t know what it measures or how precisely |
| The number of questions tracked | 500 questions nobody asks are worth less than 30 real ones |
| “AI impressions” from a third-party tool | No assistant publishes its answer volumes per business |
The minimum dashboard
- Overall mention rate, with its margin of error, week by week.
- The same, per assistant: the gap shows where to act.
- The topics (service × city) where demand is high and your rate is low: these are your opportunities.
- Visits and enquiries from AI assistants (how to measure them).
That is exactly what Recoia shows, with no single unexplained score: every number is linked to the raw answers that produced it.
Updated 26 September 2026.