How to Measure AI Visibility: A Mention Is Not a Citation or a Recommendation
AI visibility cannot be reliably expressed with a single, manually typed prompt. You need to track multiple types of prompts, platforms, and repetitions. At the same time, distinguish whether the system merely knows the brand, mentioned it, cited the website with a link, or actively recommended it.
Define the events you measure
| Event | Definition | Example |
|---|---|---|
| Entity knowledge | The system accurately describes the brand on a direct query | "What is SEO Radar?" |
| Mention | The name appears in the answer | The company is included in a list |
| Citation | The answer links to a specific web page | A source link under a claim |
| Recommendation | The system actively suggests the brand as a suitable choice | "Consider company X…" |
| Accuracy | The description of the brand and facts matches reality | Correct services, location, price |
| Sentiment | The context of the mention is positive, neutral, or negative | A recommendation vs. a warning |
An aggregate score can be useful, but it must be possible to view the raw metrics as well.
Build a query map
Split queries by stage:
Branded
- What is brand X?
- What services does X offer?
- Is X suitable for Y?
Informational
- What is a GEO audit?
- How do I allow a website in ChatGPT Search?
Comparative
- What's the difference between X and Y?
- Which tool can do A and B?
Recommendation-seeking
- Recommend a Czech tool for a technical SEO audit.
- Which company handles AI visibility for e-commerce sites?
Local and industry-specific
- SEO consultant for an e-commerce store in Prague.
- A tool for a Czech B2B website.
Prompts should be based on real customer decision-making, not just the brand name.
Record the test context
For each run, store:
- the exact prompt wording;
- the date and time;
- the platform and model, if known;
- whether web search was enabled;
- the language and locale;
- logged-in or anonymous mode;
- the position in the conversation;
- the full answer and sources;
- the repetition number.
Without this data, you can't later reproduce the result or explain a change.
Why repetition is necessary
Generative systems are non-deterministic. The same prompt can return different names and sources. That's why you should measure a rate:
mention_rate = number of answers with a mention / number of all answers
Similarly:
citation_rate = number of answers with a link / number of all answers
recommendation_rate = number of active recommendations / number of all answers
The minimum number of repetitions depends on the purpose. Smaller samples may be enough for operational monitoring, but public claims and comparisons require a more robust methodology and a confidence interval.
Don't create false precision
A score of 73/100 looks precise, but without published weights it explains nothing. State:
- which categories feed into the score;
- what weight each has;
- how missing data is handled;
- how competitors are normalized;
- how many prompts and runs were used;
- how much variability there is.
Alongside the score, always show the actual answers and a breakdown of the metrics.
Measure competitors the same way
For a fair comparison:
- use the same prompts;
- the same time period;
- the same models and modes;
- the same number of repetitions;
- the same definitions of a mention and a citation.
Don't pick competitors after the fact based on who showed up in the answer. Determine your direct business competitors in advance, and separately track newly discovered brands.
Connect AI visibility to web analytics
Track:
- referral traffic from AI services;
- UTM parameters;
- landing pages;
- conversion rate;
- assisted conversions;
- queries for which a citation occurred.
OpenAI states that it automatically appends the parameter utm_source=chatgpt.com to links from ChatGPT Search. Verify that it's preserved through redirects and in your analytics.
Recommended report structure
- Scope and time period.
- Platforms and models.
- Set of prompts by intent.
- Number of runs.
- Mention, citation, and recommendation rate.
- Accuracy and sentiment.
- Most frequently cited URLs.
- Competitor comparison.
- Change versus the previous period.
- Recommendations by impact and effort.
- Limitations of the measurement.
Common mistakes
- using a single prompt;
- using only branded queries;
- confusing a link with a recommendation;
- not recording the model and date;
- mixing answers with and without web search;
- no repetitions;
- changing prompts between periods;
- presenting a proprietary score without a methodology;
- ignoring incorrect or negative context.