AI mentions, citations and recommendations: how to read GEO performance

Key takeaways

A brand mention, a source citation and a brand recommendation are three different outcomes.

Each metric needs fixed prompts, engines, languages and run conditions, and it has to state its numerator, denominator and sample size.

Visibility can point you toward what to fix, but it is not a stand-in for audience reach, inquiries or revenue.

Picture a B2B exporter that notices AI engines often cite its spec pages yet rarely list it as a supplier option. More citations is good news, but it doesn't necessarily make buyers more likely to put the company on their shortlist. It also works the other way: when a brand is recommended, the source may be an industry publication rather than its own website.

These are hypothetical scenarios. When you assess a real project, keep the full answers and the cited URLs, then decide which pages and information need work.

What do mentions, citations and recommendations each tell you?

A mention means the answer names the brand. A citation means the answer links to a source. A recommendation means the answer lists the brand as a suitable option for a specific need. Track the three signals separately so you don't mistake an engine using your content for an endorsement.

Mentions: does AI include the brand in its answer?

Counting isn't enough; check that the description is accurate too. If the brand shows up but its products, service regions or fit criteria are wrong, that isn't simply a visibility gain.

Citations: whose information did AI use?

A citation can point to the official website, or to media coverage, industry databases or reviews. Separate the official domain from third-party sources that mention the brand, and keep a record of the actual pages.

Recommendations: does the brand meet the conditions in the prompt?

Being listed as an option, getting a positive suggestion and receiving a conditional recommendation can each be judged by different rules. Define your criteria up front; picking the favorable answers after the fact doesn't count. Record ranking separately as well, and don't assume list order means the same thing in every engine.

How do you turn results into comparable metrics?

The definitions below are the simplified ones used in this article. A real report should include full definitions, because metrics with the same name in different tools may use different denominators.

Mention rate: the number of answers that mention the brand, divided by the number of valid answers included in the comparison.

Official citation rate: the number of answers that cite the brand's official domain, divided by the number of valid answers. If third-party coverage of the brand is counted, report it as a separate metric.

Recommendation rate: the number of answers that meet predefined recommendation criteria, divided by the number of valid answers.

Citation share of voice (Citation SOV): using a fixed set of comparison brands and fixed counting rules, a brand's official citation events divided by the official citation events of all compared brands. Disclose the brand set, the domains and how duplicate citations are counted.

Competitor share: this can be calculated from mentions, recommendations or citations. The report must state which signal it uses rather than leaving an undefined "competitor share" on the page.

A worked example

Say you run 20 prompts 3 times each and get 60 valid answers. Of those, 18 mention the brand, 9 cite its website and 6 meet the recommendation criteria. That gives a mention rate of 30%, an official citation rate of 15% and a recommendation rate of 10%. These figures are for illustration only; they are not results from donguri clients.

Now suppose that across the same answers, the brand and a fixed set of competitors earn 30 official citation events in total, and the brand accounts for 9 of them. Under that definition, its Citation SOV is 30%. That figure answers a different question from the 15% official citation rate above, and the two shouldn't be used interchangeably.

You also need to decide in advance how to handle prompts that don't trigger an answer, system failures and answers with no citations. Quietly excluding different data each period and then comparing percentages doesn't work.

How should B2B exporters break down their reports?

Split results by market and language first, then by buyer task: discovering suppliers, comparing options, confirming implementation requirements and building a shortlist. This shows when a brand performs fine on general concept prompts but disappears from the prompts closest to a purchase.

A combined multi-market score works as a summary, but it shouldn't hide a gap in any single market. If you combine markets, disclose the weighting. Adding prompts or changing the market mix can also shift the overall score.

The same metric shift can have different causes

The brand isn't mentioned: first confirm the prompt is relevant to the product, then compare competitors and sources.

The brand is cited but not recommended: check whether the answer draws on specs, general knowledge or comparative evidence, and whether the brand actually meets the conditions in the prompt.

The brand is mentioned but described incorrectly: compare the official website, third-party pages and the original answer to pinpoint the conflicting facts.

Scores rise after an improvement: log what changed and when it was published, then check repeat results for the same prompts and any model changes over the same period. Correlation before and after a change does not prove causation.

How do you avoid mistaking noise for progress?

Hold the prompts, markets, languages, engines, run settings and number of repetitions constant, and keep the dates and raw answers. Record the model version as well when it's available. When the prompt set is updated, create a new version and keep the old prompts running for an overlap period so the two can be compared.

Reports should include sample size, the share of failed or missing runs, and how results changed. A single run, or only the best of several runs, isn't enough to show a sustained trend.

How does visibility connect to business results?

A position in an AI answer is not the number of people who saw it, nor is it a probability of closing a deal. Companies can pair it with branded search, identifiable AI referrals, inquiry sources, inquiry quality and sales pipeline progress, while accepting that zero-click and cross-device behavior may not be fully trackable.

donguri AI supports improvement through monitoring and diagnosis, then validates the results under consistent conditions. The value lies in mapping each gap to a piece of work, not in collapsing every result into one attractive overall score.

Next step: check how your monitoring is defined

If you already have a GEO report but aren't sure what to change next, book a donguri AI diagnostic or monitoring demo. We'll go through your prompt set, denominators, citation sources and competitor benchmarks, and look at how the report connects to content and technical improvements.

Methodology

The metrics in this article are practical measurement definitions and teaching examples, not a cross-platform standard; all example figures are hypothetical. A formal project should disclose its prompt set scope, judging rules, sample, test conditions and limitations.