Metric definitions
Numerators, denominators and confidence intervals for the organic mention, organic recommendation and risk mention rates.
On this page
Every rate in PeakGEO can be traced back to specific answers: which answers form the numerator, which form the denominator, and which are left out all follow fixed rules. This page explains how the core metrics in the console are calculated, and how to tell whether a number is solid enough to quote.
The unit of measurement: one answer
The unit is one answer produced by one model for one question.
- Asking the same question once on 4 models gives 4 answers.
- When a monitoring plan uses repeated sampling, the same question is asked several times per model per run, and each attempt counts as a separate answer.
- Related question counts are de-duplicated by question; answer counts are never treated as question counts.
Valid answers
Only valid answers enter the numerator and denominator of a rate. These are excluded:
- sampling that timed out or failed;
- model channels that are not configured or temporarily unavailable;
- empty answers returned by the model.
This keeps "we didn't get an answer" from being counted as "the brand wasn't mentioned". Failed parts can be retried from the run history, and recovered answers are counted again.
Which question types count
Only organic questions — questions that don't name the brand — feed the core metrics:
| Question type | Denominator of organic mention rate and risk mention rate | Denominator of organic recommendation rate | Numerator of organic recommendation rate |
|---|---|---|---|
| Organic discovery | Yes | Yes | No, even if the answer sounds like a recommendation |
| Organic recommendation | Yes | Yes | Yes, when the answer clearly recommends your brand |
| Brand verification (names the brand) | No | No | No |
| Custom question | No | No | No |
When a question already names the brand, the answer mentioning it is a given and says nothing about whether users would discover the brand on their own. Brand verification questions are used to see how AI describes and judges the brand, not to compute mention or recommendation rates. See Prompts.
Core metrics
| Metric | Formula | Notes |
|---|---|---|
| Organic mention rate | Valid organic answers that mention your brand ÷ all valid organic answers | Each answer counts at most once |
| Organic recommendation rate | Valid answers to organic recommendation questions that clearly recommend your brand ÷ all valid organic answers | The numerator comes only from organic recommendation questions; the denominator also includes organic discovery |
| Risk mention rate | Valid organic answers carrying a risk signal about your brand ÷ all valid organic answers | One answer with several risk signals still counts once; lower is better |
Mention vs. recommendation: "Options on the market include A, B and C" is a mention. "Given your needs, A is the one to consider first" is a clear recommendation.
Why the recommendation denominator includes organic discovery: it gives mention, recommendation and risk the same denominator, so the three can be read side by side. The overview, metric details, model and topic breakdowns in Visibility, competitor recommendation rates and exported reports all use this denominator.
The order in which brands appear in an answer is not a recommendation ranking. PeakGEO looks at whether the brand is mentioned and whether it is clearly recommended, not at answer order.
Example 1
This run has 10 valid organic answers. 4 of them come from organic recommendation questions, and 2 of those 4 clearly recommend your brand:
- Organic recommendation rate = 2 ÷ 10 = 20%, not 2 ÷ 4 = 50%.
- A recommendation-like phrase inside an organic discovery answer does not add to the numerator.
- If one answer carries two risk signals, the risk mention rate is still 1 ÷ 10 = 10%.
Example 2
This run has 4 organic discovery, 3 organic recommendation and 1 brand verification answer, all valid. Of the 7 organic answers, 5 mention your brand, 2 organic recommendation answers recommend it, and 1 carries a risk signal:
| Metric | Calculation | Result |
|---|---|---|
| Organic mention rate | 5 ÷ 7 | 71.4% |
| Organic recommendation rate | 2 ÷ 7 | 28.6% |
| Risk mention rate | 1 ÷ 7 | 14.3% |
The brand verification answer is not part of these three. Citation metrics work differently — their denominator is all 8 valid answers. See Cited sources and share of voice.
Confidence intervals and sample size
AI answers to the same question naturally vary, so a single percentage is only an estimate. PeakGEO attaches a 95% confidence interval (Wilson method) to every rate: given the current number of answers, this is the range the true level most likely falls in.
- More answers give a narrower interval and a steadier number.
- With fewer than 20 valid answers the interval is wide, so the rate shouldn't be quoted on its own or used to draw conclusions.
- Rates are shown to one decimal place.
For example, 2 recommendations out of 10 answers is a 20% point estimate, but the interval runs from about 5.7% to 51.0%. The same 20% from 100 answers (20 recommendations) narrows to about 13.3%–28.9%. When comparing two rates, PeakGEO first checks whether their intervals overlap instead of just comparing the point estimates — see Comparing runs.
"No valid answers" vs. 0%
When the denominator is zero, the console shows "No valid answers" instead of 0%.
- 0%: there are valid answers, but none mention (or recommend) the brand. That is a real measurement.
- No valid answers: there is nothing to count — for example all sampling failed, or the current filter contains no organic questions. It means "not measured", not "scored zero".
Measurement state
Each run gets an overall state based on how many question × model combinations succeeded:
| State | Condition | How to read it |
|---|---|---|
| Unmeasured | No run yet, or no combination succeeded | No usable data |
| Insufficient sample | Fewer than 5 successful combinations, or a success rate below 50% | Treat numbers as indicative only; retry or widen the scope first |
| Partial | Some combinations failed or the plan wasn't fully sampled | Metrics reflect only the valid answers received |
| Measured | Every planned combination succeeded | Read normally |
For partial runs the console notes: "Some samples failed; metrics represent valid answers only."
Reading tips
- A high mention rate doesn't mean the brand is recommended — only that AI remembers it.
- The Tracking table is a question-by-question troubleshooting tool and may include brand verification questions; its mention figures don't replace the organic mention rate on the overview.
- The overview shows the latest run, while Visibility can be filtered across multiple runs. Numbers from different scopes aren't interchangeable.
- Metrics reflect model API answers collected under uniform conditions and may differ from what an individual sees in an app.
Related: Cited sources and share of voice · Comparing runs · Glossary