AI citations, mentions and recommendations: what should you measure?

A citation, a brand mention and a recommendation are three different outcomes. Combining them into one "AI visibility" number can make progress look stronger than it is. Measure each outcome separately and pair visibility with factual accuracy.

Citation: your page was used as evidence

A citation means the answer engine exposed your page as a source. That can be valuable even when the company is not recommended: a research article, definition or dataset may influence the answer and send qualified referral traffic. But a citation alone does not prove the assistant considers your product a suitable option.

Mention: the company appears in the answer

A mention means the brand or company name is present. Inspect the context before celebrating. The answer could describe you as an alternative, repeat an outdated fact, or explicitly say you do not meet the buyer requirement. Positive, neutral and negative mentions should not receive the same score.

Recommendation: the company is proposed as a fit

A recommendation is closer to commercial intent: the assistant presents your company as an option for the stated need. Relevance matters. Being recommended to an ineligible customer or in a country you do not serve creates confusion rather than value.

A practical AI visibility scorecard

• Mention rate: the percentage of relevant prompts where the company name appears.
• Recommendation rate: the percentage where the company is proposed as a suitable option.
• Citation rate: the percentage with a visible citation to your site or a credible third-party source about you.
• Factual accuracy: the percentage of tested factual claims that are correct and current.
• Share of voice: your mentions compared with the relevant alternatives appearing for the same prompt set.
• Fit: whether the mention or recommendation matches the country, customer type and requirements in the question.

Keep the test comparable

Store the exact prompt, model, date, web-search mode and country context with every answer. Generated responses vary between runs, and model behavior changes over time. Use several observations and a stable prompt set before calling a movement a trend.

1. Separate discovery prompts from direct brand questions.
2. Score the full answer, not a screenshot cropped around the brand name.
3. Open citations and verify that they support the claim attributed to them.
4. Flag incorrect facts even when the overall tone is positive.
5. Compare with a relevant competitor set rather than every company named in the category.