How to Track Brand Mentions in AI Answers – Method and Metrics
Learn how to measure whether AI assistants mention your brand: prompt sets, sampling, what to record, and why single checks are meaningless.
Answer engines do not return a ranked list you can scrape a position from. They return prose that varies between runs, so brand visibility has to be measured statistically rather than looked up.
1. What Is Actually Being Measured
Three distinct questions get conflated, and they need different measurements.
| Question | Measure |
|---|---|
| Are we named at all? | Mention rate across repeated runs of a prompt |
| Are we cited as a source? | Citation rate — a link to your domain, not just the name |
| How are we described? | Sentiment and attribute extraction from the surrounding sentence |
A brand can be mentioned without being cited, and cited without being recommended. Reporting one number hides the difference.
2. Build a Prompt Set, Not a Keyword List
The unit of measurement is a question a buyer would actually ask, not a keyword.
- Category prompts — "best tools for X" — measure whether you are in the consideration set.
- Comparison prompts — "X vs Y" — measure how you are positioned against a named rival.
- Problem prompts — "how do I solve X" — measure whether you are the recommended mechanism.
- Brand prompts — "what is X" — measure whether the description is accurate.
Freeze the wording. Changing a prompt changes the measurement, and drift makes trends meaningless.
3. Sample, Because Answers Vary
The same prompt produces different answers on different runs. A single check tells you almost nothing, and two runs that disagree tell you the model is stochastic rather than that something changed.
Run each prompt several times per collection cycle, record every response, and report a rate with its spread rather than a yes or no. Increase repetitions when a rate sits near an interesting threshold, because that is where sampling noise is most likely to mislead.
4. What to Record Per Response
prompt id · engine · timestamp · full response text
brands named (yours and others) · citation URLs
position of first mention · sentence containing the mention
Storing the full text matters. Rates can be recomputed later with a better extractor; a stored boolean cannot.
5. Collecting the Responses
Answer engines are interactive web surfaces, so collection needs a browser session rather than an API call in most cases. Practical requirements:
- A consistent geography and language, since answers vary by both.
- A clean session per run, so conversation history does not contaminate the sample.
- Recorded engine version or date, because model updates shift results sharply.
- Storage of the raw response before any parsing.
6. Turning Responses Into Metrics
- Mention rate — runs naming the brand ÷ total runs, per prompt and aggregated.
- Citation rate — runs linking the domain ÷ total runs.
- Share of voice — your mentions ÷ all brand mentions in the same responses.
- Descriptive accuracy — share of mentions where the stated capability is correct. This one is manual and worth the effort, because a confident wrong description is worse than absence.
7. Common Mistakes
- Treating one run as a result. Variance between identical runs is often larger than the change being investigated.
- Measuring only your own brand. Without competitor mentions there is no denominator and no context.
- Ignoring citation versus mention. They respond to different work: being named follows from reputation, being cited follows from being crawlable and quotable.
- Changing prompts mid-series. Comparability disappears the moment wording drifts.
8. Summary
Track brand presence in AI answers by running a frozen prompt set repeatedly, storing whole responses, and reporting mention, citation, and share-of-voice rates with their variance. Answers are probabilistic, so the measurement has to be too.
The output that changes decisions is rarely the headline rate. It is the sentence around the mention — what the engine thinks you do, and whether that is accurate.