Measure GEO performance as four separate rates — mentioned, accurate, recommended and cited — each calculated over a stated number of valid answers. Collapsing them into one "AI visibility score" makes every change impossible to act on: mentions can rise while accuracy falls, and a single number reports that as progress. GEO (Generative Engine Optimization) is the work of getting a brand found, described correctly and recommended in AI-generated answers; this page defines the metrics used to measure it. For the step-by-step workflow of collecting answers, see how to measure brand visibility in ChatGPT.
The four metrics
| Metric | Definition | Formula |
|---|---|---|
| Mention rate | The answer names the brand at all | answers mentioning the brand ÷ valid answers |
| Accuracy rate | What the answer says about the brand matches confirmed facts (product, market, price range, capabilities) | accurate mentions ÷ answers mentioning the brand |
| Recommendation rate | The brand is in the set the user is told to choose from, or is the suggested choice | answers shortlisting or recommending the brand ÷ valid answers |
| Citation rate | The answer uses or links a page you control or a verified source about you | answers citing such a page ÷ valid answers (retrieval track only) |
Two details that change results:
- Accuracy uses mentions as its denominator, not all answers. A brand that is rarely mentioned but always described correctly has a high accuracy rate and a low mention rate — two different problems with two different fixes.
- Mentioned is not recommended, and recommended is not accurate. An answer can list you as an option while describing the wrong product.
Denominators and when not to publish a number
A rate without a denominator is not a measurement. Every report should state:
- how many questions, split by type (category, comparison, branded);
- how many runs per question;
- which models, with web search on or off;
- the time window;
- how many answers failed and were excluded.
Valid answers are the denominator; failed, refused or unusable answers are counted separately and never silently dropped. When the base is too thin — a model that returned four usable answers, for example — the honest output is the raw count with a note, not a percentage. "1 of 4" and "25%" describe the same data, but only one makes the small sample obvious.
Keep the tracks separate
Report these as separate tracks rather than averaging them:
- Category vs branded questions. Category questions measure whether you are found; branded questions measure whether you are described correctly. Mixing them inflates the mention rate.
- Retrieval (web search on) vs closed-book. With search on, answers depend on pages the model can retrieve this week; closed-book answers reflect what the model learned in training and change only with new model versions. Citation rate exists only on the retrieval track.
- Model by model. Answer length and listing habits differ by model, so compare each model with itself over time before comparing models with each other. Record how each model was accessed: models called through the same gateway can share retrieval results and should not be treated as independent samples.
Reading a change
Before calling a change a result, check that the question set, its version, the models, the search setting and the number of runs are the same as last time. If anything changed, mark the comparison as broken. With small samples, a one-round move is a direction to watch, not a conclusion; look for the same movement across two or more rounds.
Also look beyond the four rates. The list of most-cited domains for your questions shows which pages the models rely on — and therefore where accurate information about you needs to exist.
Two examples from real measurements
Both rules above come from mistakes that are easy to make with real data.
Small samples per model. According to our measurement on 26 September 2026 (17 buyer-style questions, 4 models with web search on, each question asked once), there were 38 valid answers in total: 15 from DeepSeek, 14 from GLM, 5 from Qwen and 4 from Kimi. A combined mention rate over 38 answers looks solid, but the per-model rates for the last two rest on four or five answers each. The right report shows the counts per model and avoids percentages where the base is that thin.
Shared access routes. According to our measurement on 22 September 2026 (22 unbranded category questions, 5 models with web search on), DeepSeek, GLM and Qwen returned identical citation counts for several sites — 7 for IT Home, 5 for ifeng.com and 2 for CSDN. The most likely reason is that those three models were called through the same cloud platform that day, so their retrieval results overlapped. Treated as three independent models, the data would have overstated how consistently those sites are used. Recording the access route for each model is what made the overlap visible.
A one-page reporting template
| Section | Contents |
|---|---|
| Scope | Questions (by type) and version, models and access route, search on/off, runs, window |
| Base | Answers collected, failed and excluded, valid |
| Rates | Mention, accuracy, recommendation, citation — per model and per question type, with counts |
| Sources | Most-cited domains; whether your pages appear |
| Changes | What changed since last round; comparability note |
| Evidence | Two or three raw answer excerpts with question ID, model and date |
FAQ
Is there a single GEO score I can track? You can compute one, but it hides which of the four things moved. Track the four rates and report a composite only alongside them.
How long before GEO work shows up in these metrics? There is no reliable public figure; it depends on crawl and index cadences that engines do not publish. Measure it: freeze a question set before the change and retest under the same conditions afterwards.
Can keyword rankings replace these metrics? No. Generative answers are not ranked lists and vary between runs, so the unit is a sampled rate, not a position.
What is a good mention rate? It depends on category and competition. Compare against your own earlier rounds and against named competitors measured on the same questions, rather than a universal benchmark.