- "AI visibility" collapses three distinct outcomes into one word: citation (a model links or names your domain as a retrieval source), mention (a model describes your brand at all, sourced or not), and recommendation (your brand appears in a competitive best-of or comparison list). AI visibility strategy names this distinction; this piece is the full reference on it.
- Every citation is also a mention. Most mentions are not citations: a model can describe your brand entirely from training data, with no retrieval step and nothing to link to.
- Citations leave a trail you can audit (server logs, retrieval activity). Mentions don't: the only way to check what a model says about you from memory is to ask it directly and read the answer.
- Recommendation is the outcome most tied to competitive, third-party proof (reviews, comparison content, press) not to how well-structured your own site is. Content extractability work moves citations; it doesn't reliably move recommendation rate on its own.
- A mention with no citation is also where stale or incorrect brand facts live longest, because there's no source attached to check the claim against, and no retrieval event to re-trigger a refresh.
AI visibility strategy makes this distinction in a single paragraph: citation, mention, and recommendation are three separate outcomes, and a tactic that moves one doesn't automatically move the others. That's the right amount of space for a strategy piece to give it, but the distinction itself is easy to gloss over in practice, especially once "AI visibility" gets reported as a single blended score. This is the full reference: precise definitions, why each one requires a different way of checking it, and which one actually matters for a given goal.
What is a citation?
A citation is the narrowest and most checkable of the three. It happens when a model's answer traces a specific claim back to your domain: a named source, a linked URL, or both, depending on the platform. What does it mean to be "cited" by an AI model? covers the full mechanics: the fetched-vs-cited gap documented in OpenAI's and Anthropic's own API schemas, why citation behavior differs by product (Perplexity cites by default, ChatGPT and Claude only when web search is actively invoked, Google AI Overviews shows a source panel that functions more like general attribution), and the Semrush finding that roughly 62% of the time a source gets used, it never gets named in the visible text at all: a "ghost citation."
The practical implication that matters here: a citation is downstream of retrieval. Something had to be fetched for something to be cited. That means citations are the one outcome in this trio with an actual audit trail: server logs showing crawler activity, API responses that return a sources or citations field, a retrieval event that either happened or didn't.
What is a mention?
A mention is broader and looser: the model says something about your brand at all, whether or not it names a source, whether or not what it retrieved (if anything) actually gets attributed. Critically, a mention can happen with zero retrieval: a model trained on data that included your brand can describe your category, your positioning, even specific claimed facts about you, purely from what it learned during training, with nothing fetched and nothing to link to.
This is the outcome most likely to carry an error nobody catches. A citation, when it happens, at least points at a checkable source: right or wrong, you can trace the claim. A mention sourced from training data has no such trail. If a model's training data included your 2024 pricing and you've since changed it, that stale number can keep surfacing in mentions indefinitely, with no retrieval event to trigger a refresh and no source attached that a fact-checker (human or automated) could flag. This is a large part of why AI visibility strategy's monitoring stage exists as ongoing work rather than a one-time check: a mention-level error doesn't announce itself the way a broken citation link would.
Every citation necessarily involves the model mentioning your brand: you can't be cited without being named. The reverse doesn't hold: the large majority of times a model discusses a brand, it's a mention with no citation attached, drawing on memory rather than a live lookup.
What is a recommendation?
Recommendation is the narrowest and most competitively loaded of the three: your brand shows up in a list of options a model surfaces for a comparison or "best of" query, positioned against named alternatives rather than discussed in isolation. A single-source factual citation ("according to [brand]'s pricing page...") carries no competitive framing at all. A recommendation always does: you were picked, or included, when the model could have named someone else instead.
This is also the outcome furthest from what on-site content work can move directly. Structuring your own pages so a crawler can extract them cleanly (content extractability) raises your odds of being cited when your page is the relevant source for a specific factual claim. It does comparatively little to raise your odds of being recommended in a "best CRM for a 10-person sales team" query, because that outcome depends far more on third-party proof the model can weigh against competitors (reviews, independent comparison content, press coverage), the off-site and earned work covered in why earned media outperforms owned content in AI citations. A vendor pitch that promises "better content structure gets you recommended more" is quietly assuming recommendation moves the same way citation does. The evidence says it doesn't.
Why does this distinction change how you measure, not just how you talk?
The three outcomes require three different measurement approaches, which is the real reason conflating them causes problems beyond vocabulary:
| Outcome | How you'd actually check it | What moves it |
|---|---|---|
| Citation | Server logs (crawler activity), API sources/citations fields, direct prompting with retrieval enabled | Access & crawlability, content extractability |
| Mention | Direct prompting only: no crawl trail exists for training-data-sourced descriptions | Training-data presence, entity consistency, historical off-site footprint |
| Recommendation | Direct prompting against named competitors on comparison-style queries specifically | Third-party proof: reviews, comparison content, earned press |
A single blended "AI visibility score" that reports one number across all three tells you less than tracking them separately: it's the same problem AI visibility score documents for blending across platforms, applied to blending across outcome types instead. A brand can have a strong citation rate (well-structured, crawlable content) and a weak recommendation rate (thin review presence) at the same time, and a blended score would report something unremarkable in the middle, hiding the exact gap that's actually actionable.
Frequently asked questions
What's the difference between a mention and a citation in AI search?
A citation is a model naming or linking your domain as the source behind a specific claim in its answer: it traces back to a retrieval step. A mention is a model describing your brand at all, correctly or not, whether or not it names a source: it can come entirely from training data, with no retrieval involved and nothing to link to. Every citation involves a mention of your brand; most mentions involve no citation at all.
Can a brand be mentioned by an AI model without being cited?
Yes, and it's the more common case. A model trained on data that included your brand can describe you, compare you to competitors, or state facts about you from memory alone, with no live retrieval and no source attached. This is also where inaccurate mentions live (stale pricing, a discontinued feature, a category you no longer compete in) repeated with no citation to correct against, because nothing was actually looked up.
How is a "recommendation" different from a citation or a mention?
A recommendation is a specific competitive outcome: your brand appears in a list of options a model surfaces for a comparison or best-of query, positioned against named alternatives. A citation can happen with zero competitive framing (a single-source factual answer); a recommendation always implies you were picked over, or alongside, competitors, which depends far more on third-party proof (reviews, comparison content, press) than on how well-structured your own pages are.
Which one should I actually be tracking?
Depends on the goal. If the question is 'can a crawler find and extract our content,' track citations: they leave a server-log and retrieval trail you can actually audit. If the question is 'does the model already know who we are, and does it know us correctly,' track mentions via direct prompting, since there's no crawl trail to check. If the goal is competitive share in AI-driven comparison shopping, track recommendation rate against named competitors specifically, not a generic visibility score that blends all three into one number.
Sources referenced
This piece is the full reference behind the citation/mention/recommendation distinction made in AI visibility strategy. For the technical mechanics of citation specifically (including the ghost-citation data referenced above), see what does it mean to be "cited" by an AI model?; for why recommendation depends on off-site proof, see why earned media outperforms owned content in AI citations. For how citation accuracy itself varies sharply by platform, see Perplexity vs. ChatGPT citations and Grok's AI citations.
