Measurement guide
GEO metrics: what to measure to know whether AI recommends you
The five core GEO metrics are mention rate, share of voice, citation rate, average position and sentiment. They are complemented by the accuracy of what AI says about you and by the traffic arriving from assistants. Because answers vary, every rate should be read with its margin of error. GeoRush calculates them by engine, question and market.
Updated on

Why GEO needs its own metrics
In SEO you measure positions in a results list that changes little from one search to the next. In GEO (generative engine optimization) you measure written answers, which can change every time someone asks. The same question to ChatGPT can name you on one try and not on the next. That is why GEO metrics are frequencies: out of a hundred answers, how many include you, in what position, with what tone and which sources?
This has two practical consequences. First, you need many answers for a figure to mean anything: several questions, several engines, several repetitions. Second, you should read each rate with its margin of error, so you don’t mistake noise for progress.
The core metrics
| Metric | Formula | What it tells you | Watch out for |
|---|---|---|---|
| Mention rate | Answers that name you ÷ total answers | How often you exist for AI on your questions. | With few answers the margin of error is wide. |
| Share of voice | Your mentions ÷ mentions of all tracked brands | Your space relative to competitors. | Depends on which competitors you track; keep the list stable to compare. |
| Citation rate | Answers linking to your site ÷ total answers | Whether AI uses your content as a source. | Some engines don’t show sources on every answer. |
| Average position | Average place you are named in, when you appear | Whether you are the first option or one of many. | Only counts answers where you appear; read it with mention rate. |
| Sentiment | % of positive, neutral and negative mentions | How it describes you. | A negative mention can be worse than not appearing. |
| Accuracy | % of claims about you that match your official facts | Whether AI says true things about you. | You need an official fact sheet to compare against. |

Worked example with a margin of error
Illustrative example (made-up figures to explain the calculation). The 95% Wilson interval is a classic statistical method for estimating a proportion from few observations:
Questions: 10 Engines: 4 Repetitions: 1 -> 40 answers
Answers that name you: 14
Mention rate = 14 / 40 = 35%
95% Wilson interval: 22% – 50%
With twice the answers (28 of 80): 35%, interval 25% – 46%
With 400 answers (140 of 400): 35%, interval 30% – 40%
Reading: with 40 answers, moving from 35% to 42% may be noise.
With 400, a rise like that is hard to explain by chance.How to read the margin of error
If this month your mention rate is 35% (interval 22–50%) and last month it was 30% (interval 18–45%), the intervals overlap heavily: you can’t claim you improved. If, with more answers, you go from 30% (26–34%) to 38% (34–42%), the improvement is far more credible. Rule of thumb: before celebrating or panicking, check whether the intervals overlap.
For the same reason, be wary of any report that gives a mention rate without saying how many answers sit behind it. GeoRush shows each rate with its 95% confidence range (Wilson interval) so you can see when a change is real.
Traffic and business metrics
Answer metrics tell you whether AI recommends you. These tell you whether that reaches your site:
| Metric | How to get it | Note |
|---|---|---|
| Visits from assistants | Web analytics, by referrer (chatgpt.com, perplexity.ai, gemini.google.com, etc.) and by campaign parameters. | According to OpenAI, ChatGPT adds utm_source=chatgpt.com to referral links. |
| Impressions in Google’s AI features | Search Console’s generative AI performance report. | Since 2026; shows impressions of your URLs in AI Overviews, AI Mode and Discover, by page, country, device and date. |
| AI crawler visits | Server logs or analytics that identify agents such as OAI-SearchBot, PerplexityBot or Claude-SearchBot. | Shows whether assistants read your pages; not the same as being cited. |
| Conversions from AI | Goals or funnels in your analytics filtered by those sources. | Volumes are often small: look at trends, not single days. |
Part of AI’s value leaves no trace in analytics: someone reads your name in an answer and later searches for you or contacts you directly. That is why answer metrics matter even when traffic is low.
How to build your GEO dashboard
- 1
Fix the question set
Between 20 and 50 buying questions grouped by topic. Changing them every month makes comparison impossible: add new ones, but keep a stable core.
- 2
Choose engines and markets
The assistants your customers use, in their language and country. A brand can be strong in Perplexity in English and absent from Gemini in Spanish.
- 3
Decide on repetitions
Repeating each question several times gives more reliable rates. More answers, narrower intervals.
- 4
Set the baseline
Measure before changing anything. Without a starting point there is no way to attribute improvements.
- 5
Set the cadence
Weekly to catch drops and competitor moves; monthly to report trends.
- 6
Break it down
By engine, topic, market and competitor. Global averages hide where the problem is.
- 7
Tie it to actions
Log what you changed and when, and compare before and after on the affected questions.
A summary score: the GEO Score
Five metrics by engine, topic and market are a lot of numbers for a leadership meeting. That is why many tools summarize them in an index. In GeoRush it is the GEO Score, from 0 to 100, combining five weighted pillars: visibility, share of voice, citation, reputation and technical readiness. It is useful for the overall trend; to decide what to do, always drill down by pillar, engine and question. Any summary index, from any tool, should break down into its source metrics.
See how this data looks in AI visibility, and how it reaches leadership or clients in reports. To connect visibility with visits, GeoRush’s web analytics separates traffic from assistants and AI crawler visits; the AI traffic page explains it in detail.
Common mistakes when reading GEO metrics
- Drawing conclusions from a screenshot or a single answer.
- Comparing months with different question sets.
- Looking only at the global average and not the breakdown by engine and topic.
- Confusing mentions with citations: you can be in the text without your site being linked, and vice versa.
- Ignoring sentiment and accuracy: appearing often while AI says false things is not success.
- Expecting lots of traffic: the value of an AI recommendation often leaves no trace in analytics.
Measure your starting point
The free check asks one engine three buying questions about your category and tells you who appears instead of you.
Sources
Official documentation and primary sources checked on November 17, 2026. These policies change often: check the source before making important decisions.
- Wilson, E. B. (1927), «Probable inference, the law of succession, and statistical inference», JASA
- OpenAI Help Center: publishers and developers FAQ
- Google Search Central Blog: Search generative AI performance reports in Search Console (June 2026)
- Google Search Central: AI features and your website
- OpenAI: overview of OpenAI crawlers
- Perplexity: Perplexity crawlers
- Anthropic: does Anthropic crawl data from the web and how can site owners block the crawler?
Frequently asked questions
Keep reading
Measure AI visibility
The method for measuring whether AI mentions you: questions, repeats, metrics and margin of error.
AI visibility
Mention rate with a 95% range, share of voice, citations and GEO Score across 9 AI engines.
Measure GEO ROI
Visibility, AI traffic and conversions: how to prove GEO works.
90-day GEO plan
A five-phase plan to measure, fix, create content, win sources and report within a quarter.