Mentions, citations, and recommendations are not interchangeable. A mention tells you that your brand entered an answer. A citation shows that a source connected to your brand helped support an answer. A recommendation means the system presented your brand as a suitable choice. The three can overlap, but each describes a different outcome.
That distinction sounds simple. In practice, it is where many AI visibility reports go wrong. One dashboard celebrates a rising mention count while the brand is still absent from buying shortlists. Another treats every citation as a vote of trust, even when the cited page is unrelated to the recommendation. A third compresses everything into one score and leaves the team unsure what to fix.
The useful question is not, “Which metric wins?” It is, “What does each metric tell us about the buyer journey, and what action follows?”
The quick definition
- Mention rate is the share of tracked answers in which the brand appears.
- Citation rate is the share of answers that include a citation to a relevant brand-owned domain or another source you are measuring.
- Recommendation rate is the share of eligible answers in which the brand appears as a recommended option, often with a position or shortlist threshold.
These definitions need a declared denominator. “Twenty mentions” means very little without knowing whether you tested 25 prompts or 2,500, which models ran, which locations were used, and whether every prompt could reasonably produce a brand recommendation.
Mentions measure awareness, not preference
A mention is the broadest visibility signal. It can show that an AI system associates your brand with a category, problem, competitor, or topic. That makes mention rate useful for checking basic retrievability and category recognition.
But a mention can be positive, neutral, or negative. Your brand might appear in a historical aside, a list of alternatives, or a warning about an unsuitable use case. It might be named after five competitors. Counting all of those as equal “wins” hides the part a buyer would care about.
Read mentions with context. Capture the prompt, answer, model, date, placement, nearby language, and competitors. Then separate category mentions from recommendation mentions. If mention rate rises while recommendation rate stays flat, the brand may be better known without becoming more persuasive.
Citations measure evidence use, not endorsement
Citations show which pages an answer surface exposed as supporting sources. OpenAI explains that public sites can appear in ChatGPT search and that allowing OAI-SearchBot helps content become discoverable, cited, and linked. Microsoft’s AI Performance reporting similarly distinguishes citation counts from ranking, authority, or a page’s role inside an individual answer.
That caveat matters. A citation is not automatically an endorsement. A model can cite your documentation while recommending someone else. It can cite a third-party review that describes your limitations. It can also cite a useful article from your domain without mentioning your product at all.
Inspect citation ownership, relevance, verification, and proximity to the claim. A verified citation to a product comparison that directly supports your differentiator is more actionable than a raw count from loosely related pages. Our documentation explains how Swep separates raw and verified citation signals so teams do not confuse detection with evidence quality.
Recommendations measure commercial inclusion
Recommendations are usually closest to the commercial outcome. If a buyer asks for the best tools for a defined job, did the answer include you? Where did you appear? What reason did it give? Which competitor occupied the position you wanted?
This metric still needs careful design. Not every prompt should recommend a vendor. An informational prompt such as “what is AI visibility?” should not be forced into the same denominator as “best AI visibility tools for SaaS.” Create an eligible prompt set and define what counts before collecting results.
For ordered answers, track position and top-three presence alongside recommendation rate. For unordered answers, record inclusion and supporting language. Do not invent precision when the answer itself does not rank the options.
Read the three metrics as a funnel
The clearest interpretation is a diagnostic funnel:
- Can the system retrieve and place the brand? Start with mentions.
- Can it find usable evidence? Inspect citations and cited pages.
- Will it choose the brand for the buyer’s need? Measure recommendations and position.
Low mentions and low citations suggest a discovery or category-clarity problem. Strong mentions with weak citations suggest the brand is known but its owned evidence is not being used. Strong citations with weak recommendations suggest the evidence is visible, but the positioning, proof, fit, or third-party consensus may not support selection. Strong recommendation rate with weak owned citations can still be fragile because the answer depends on sources you do not control.
Add segmentation before adding another score
An overall average can conceal the decisions you need to make. Segment results by prompt intent, product, audience, market, model, and competitor set. A company may perform well for category education but disappear from comparison prompts. It may win for small teams and lose for enterprise requirements. Those are different content and proof problems.
Your buyer-intent prompt library should therefore be stable enough to compare over time and specific enough to diagnose. Keep a core benchmark set, then add an exploration set for new questions. Do not quietly replace losing prompts with easier ones.
A weekly measurement routine
- Review changes in mention, citation, and recommendation rates against the same core prompt set.
- Open the underlying answers for the largest gains and losses.
- Group gaps by intent and competitor, not only by model.
- Inspect the pages and third-party sources supporting the winning answers.
- Assign one evidence, content, or technical action with an owner.
- Record the change and wait for enough repeated observations before claiming impact.
AI answers vary. One run is evidence of an observation, not proof of a durable trend. Use repeated measurements, preserve snapshots, and state sample sizes. If a metric moves after a page update, treat causation as a hypothesis until the pattern holds across time and relevant prompts.
How Swep uses the metrics
Swep brings prompts, answers, competitors, mentions, rankings, and citations into one workflow. The purpose is not to crown a universal metric. It is to connect an observed outcome to the evidence behind it and the next page or proof asset worth improving.
Start with the business question. If the problem is category awareness, mentions matter first. If the problem is missing source support, investigate citations. If the problem is losing shortlists, recommendations and competitor reasoning deserve priority. Then use the other metrics to explain why.
The metric that matters is the one that changes a decision without hiding the evidence. Everything else is decoration.