Mention, Citation, Recommendation: Three Wins, Three Different Plays

"We got mentioned by ChatGPT" can mean three completely different events. Until you can tell them apart, you cannot tell which of your efforts is working, or why a competitor with worse content keeps winning the deal.

What is the difference between a mention, a citation and a recommendation?

A mention is your name appearing inside the generated text itself. The model wrote your name because it knows it, usually from training data, sometimes from a retrieved source it read but did not link. A citation is a linked source attached to the answer: the engine retrieved your page, leaned on it while composing, and credits it with a footnote or source card. A recommendation is the engine advising the user to choose you: "for this problem, talk to this person." It may arrive with or without a link. It is the only one of the three that functions as a referral rather than a reference.

These get collapsed into one metric by most dashboards, which is how teams end up celebrating citation counts while a rival owns the recommendation on every question that pays. The distinction is not academic. It decides where your next quarter of effort should go.

The three-tier framework

• Tier 1, Mention: produced by model memory · signals familiarity · invisible in analytics · earned by widespread, consistent coverage · compounds slowly, decays slowly
• Tier 2, Citation: produced by live retrieval · signals relevance of a specific passage · visible as referral traffic · earned by retrievable, liftable content · arrives fast, churns fast
• Tier 3, Recommendation: produced by entity confidence, memory plus retrieval agreeing · signals trust · visible as inbound that says "the AI told me about you" · earned by owning a narrow question end to end · slowest to build, hardest to displace

Tier one: the mention, familiarity in the weights

When a model names you without retrieving anything, it is drawing on patterns absorbed during training: your name co-occurring with a topic across many independent sources. This is why unlinked press, podcast transcripts, directory listings, and community threads matter even though no analytics tool will ever attribute a visit to them. They teach the next model generation that your name and your specialty belong in the same sentence.

Co-occurrence is the operative mechanism. A model that has seen "Jane Ito" near "pricing strategy for marketplaces" across a podcast transcript, a conference speaker page, two community threads, and a trade publication does not need to retrieve anything to volunteer her name when the topic arises. No single one of those surfaces would have been worth much on its own, and none of them shows up in her analytics. Together they wrote her into the model's sense of the topic.

The play for mentions is breadth and consistency. Same name, same positioning, across as many crawled surfaces as you can legitimately reach. It is slow, and that is the point: mentions persist even on days when retrieval returns nothing of yours, and they bias what the engine does with whatever it retrieves. The mechanics of how engines weigh that accumulated evidence are covered in how AI decides who to recommend.

Tier two: the citation, winning the retrieval round

Citations are a retrieval event. The engine searched, your passage scored, the model used it, and the interface linked it. Because they are linked, citations are the tier you can measure directly through referral traffic in your own analytics, unlike mentions, which leave no trail at all. Anecdotally, teams that start tracking AI referrals for the first time tend to notice two things: the volume is small next to organic search, and the visitors who do arrive tend to already understand the problem, because an engine has already framed it for them before they clicked through.

The play for citations is passage-level engineering: self-contained sections, question-shaped headings, specific verifiable facts, on pages the AI crawlers can reach. The catch is durability. Citations are recomputed with every answer, which means the set of pages an engine leans on for a given question can shift as the engine re-crawls, as competitors publish, and as the model itself is updated. A citation is rented, not owned, which is why we treat the rotation problem as its own discipline, and why passage-level work covered in chunk theory and passage-level optimization matters more for this tier than page-level polish does.

Tier three: the recommendation, the conversion event

A recommendation happens when the engine is confident enough to put its own credibility behind a name. Mechanically, that confidence tends to appear when the model's memory and its retrieved sources agree: the name it already associates with the problem is also the name showing up in fresh, corroborating results. One strong page cannot produce that. An entity can.

The play for recommendations is narrowness plus corroboration. Pick the specific question you want to own, the kind of high-intent query we map in money queries, and make every layer agree about you on it: your own canonical pages, third-party coverage, consistent bios, structured identity. The engine should find no seam between what it remembers about you and what it retrieves about you. When that alignment holds, you stop appearing as a source at the bottom of the answer and start appearing as the answer.

What that looks like in practice: a fractional CFO who wants the recommendation on "who should a bootstrapped SaaS company hire for finance" needs her own site to answer that exact question in liftable passages, needs two or three independent surfaces, a podcast appearance, a finance community thread, a directory listing, repeating the same positioning about the same name, and needs her bios to agree on all of it. None of those pieces is impressive alone. The recommendation emerges from their agreement, because agreement across sources the engine cannot see as coordinated is the closest thing a machine has to trust.

Where does each tier show up, engine by engine?

The tiers are weighted differently across the major surfaces, which matters when you decide where to look for progress. ChatGPT is the mention-and-recommendation heavyweight: it answers conversationally, often from memory, and only attaches citations when its search layer is triggered, so a strong entity can be named there without a single link appearing. Perplexity sits at the opposite pole. Its interface is citation-first, nearly every claim carries a footnote, and its retrieval behavior is documented openly in its developer documentation, which makes it the easiest engine to earn and audit citations on. Google's AI Overviews fall in between: source cards reward retrievable passages, while the surrounding synthesis follows Google's existing understanding of entities. If your citations rise on Perplexity while ChatGPT still never says your name, you have a retrieval win and a familiarity gap, and the framework tells you exactly which play to run next.

Which tier should you chase first?

In practice the sequence runs backwards from how most people work:

1. Start with citations, because they respond fastest. Retrievable, passage-optimized pages can begin earning citations within weeks of being crawled, and they generate the measurable feedback you need.
2. Build mentions in parallel, because they are slow and you cannot rush them later. Every guest appearance, quoted comment, and directory entry is a deposit into the next training run.
3. Aim everything at one recommendation, on one narrow question. Owning "who should a Series A founder hire for positioning" outperforms being footnoted on fifty broad ones.

The mistake to avoid is treating the tiers as a funnel you passively ascend. Citations do not automatically mature into recommendations. They only do when they accumulate around a single unambiguous entity, which is why identity consolidation sits underneath all three plays.

The other common failure is publishing under a brand that hides the person. Anonymous company blogs can earn citations forever without ever generating a mention of a human being, because there is no name in the passages for the model to learn. If the goal is a personal recommendation, the byline, the bio block, and the claims all have to carry the same human name, on your site and in every third-party appearance. Content that wins Tier 2 while starving Tiers 1 and 3 is the most expensive kind of success in this discipline.

How do you measure each tier?

How does an engine actually decide whether to cite you at all?

Underneath the interface, a citation is the output of a retrieval step that happens before generation starts. The engine takes the user's question, converts it into a vector representation, and compares that vector against an index of passages it can search, some drawn from a live web crawl, some from a curated corpus. Passages whose vectors sit close to the question's vector get pulled into the model's context window, and the model then writes an answer grounded in whatever it was handed. If your passage never gets pulled, it does not matter how well-written it is, because the model never saw it for that particular question. This is why citation work lives or dies on retrieval, not on eloquence, and why the mechanics of chunking a page into passages a retriever can actually match are covered separately in chunk theory and passage-level optimization.

Two implications follow directly from that mechanism. First, a page can be factually excellent and still never get cited if its sentences are phrased so differently from how people ask the question that the vectors never land close together, which is the core argument behind answer-shaped writing. Second, being retrieved is necessary but not sufficient. The model can pull your passage into context and still choose not to lean on it, or lean on it without surfacing a link, which folds back into the mention-versus-citation distinction: retrieval without an attached source card is functionally a mention that happened to be triggered by search rather than training.

A worked example: the same question, three different outcomes

Consider a hypothetical fractional CFO named Renata Alves, who works exclusively with bootstrapped SaaS companies preparing for their first debt raise. Picture the same question, "who should a bootstrapped SaaS founder talk to about debt financing," put to an engine on three separate occasions as her visibility work matures.

A hypothetical progression showing how the same question can surface a mention, a citation, or a recommendation depending on what has accumulated around the name.

Nothing about Renata's underlying expertise changed between the three stages. What changed was the evidence available to the engine: first a thin trace in training data, then a retrievable page, then a corroborated entity. The lesson generalizes past this one hypothetical: the tier you get is a function of accumulated evidence, not a function of how good your best page is on any given day.

Comparing the three tiers at a glance

The three tiers side by side, showing why they need three different kinds of evidence to earn.

Why does anonymous or brand-only content struggle to earn any of the three tiers for a person?

A company blog with no bylines can absolutely earn citations. Its pages can be well-structured, well-chunked, and frequently retrieved. What it cannot do is generate a mention or a recommendation of a specific human, because there is no name in the passages for the model to learn, and no entity for memory and retrieval to agree on. The organization gets credited, not a person. For a personal brand, this is the gap that quietly caps growth: a founder or consultant can pour years into a company blog and still be unrecognized by name, because every passage the engine retrieved said "our team" or "the company" instead of a name it could carry forward into the next answer.

The fix is not abandoning brand content. It is making sure the person's name appears as the author, in the bio, and ideally in a first-person or third-person definitional sentence somewhere on the page, the same discipline covered for an individual's own site in the machine-readable About page teardown. A byline is not decoration here. It is the only mechanism by which retrieval-driven citations for a company have any chance of compounding into mentions and recommendations for the person behind the work.

Citations show up in your analytics as referral sessions from AI domains, and in the source cards of the answers themselves. Mentions require asking: run a fixed set of prompts monthly and record every time your name appears in generated text, linked or not. Recommendations require reading the answer's framing: is the engine listing you among sources, or telling the user to contact you? Record the verbatim wording, not just a yes or no, because the framing is the data: "one option is" and "the person most often recommended for this is" are different tiers wearing the same name. Log all three separately, because they trend independently and respond to different work. The full measurement routine, prompt sets, cadence, and what to record, is laid out in how to track your AI visibility. If you want the baseline run for you, with all three tiers separated on the questions that matter in your market, that is the starting point of our services.