Cited Isn't Named: Why AI Quotes Your Site but Recommends Someone Else

An AI answer is built in two moves: the engine retrieves a set of sources, then it synthesises a recommendation from everything it has read. Being cited means your page made it into the first move — it was reachable, relevant to the query, and pulled in as a source. Being named means you made it into the second — the synthesis actually landed on your brand as something the reader should consider. They are not the same event, and one does not imply the other. Your pricing page can be cited for a fact while the answer recommends the competitor that everyone else writes about. On one client account this month, the domain was cited as a source in 83 answers and named as a recommendation in none of the unbranded ones — the engine was reading the site diligently and recommending someone else. A tool that collapses this into a single "visibility" score cannot tell you which of the two is happening.

Because they have different causes. A citation is mostly a plumbing outcome: the engine could reach your page, parse it, and judged it relevant, so it pulled it in. Being named is a reputation outcome: across everything the engine read — not just your own site — the consensus pointed at you. You can win the first and lose the second when the model reads your page for a detail but recommends the brand that dominates the third-party sources it trusts for that question. The reverse happens too. A well-known brand is often named straight from the model's training and the wider web, with no live citation of its own site at all. That is a real win, but a fragile one — it rests on the model's memory rather than on anything you control, and it can move on the next model update. Neither pattern is visible if you only track one number, and both are noisy enough that you should read them across several runs, not one (we wrote about why a single check misleads in Ask Five AI Engines, Get Five Answers).

Both, but they answer to different people. Named is the commercial signal — the AI-era shortlist. It tracks whether a buyer asking an unbranded question actually hears your name, and its share against competitors. That is the number a client wants to see move. Cited is the supply signal. It tells you the engine can reach and use your content at all, which is the precondition for being named from live retrieval rather than from stale memory. If you track only named, you cannot tell a reputation problem from a crawlability one — the score is low either way. If you track only cited, you can celebrate being read while the answer quietly sends the customer elsewhere. The useful information is in the two side by side, and specifically in the gap between them.

In CrunchJunkie the two live on one screen on purpose. On the Overview, Visibility and Share of Voice measure named — how often the answer recommends you, and your slice of all brand recommendations. On Source performance, Citations, Source appearances and the citation rate measure cited — how often your domain is pulled in as a source. They are sampled the same way, several runs per prompt with a margin of error, so the two are directly comparable rather than two instruments you have to reconcile by eye. Read the gap. High citation rate with low named is a consensus problem: the engine reads you and still doesn't rate you. Named holding up while citations sit near zero is the fragile-memory case: you are being recommended from training data, not from your live site. Same brand, same week, two very different to-do lists — which is the whole reason to keep the measurements apart.

This is a consensus problem, not a plumbing one. The engine can already read you; it just doesn't conclude that you are the answer. The fixes are the unglamorous reputation ones. Get named on the third-party pages the engine already trusts for that query — the roundups, comparisons and community threads it keeps citing — because those are where the recommendation is actually being formed. Earn independent reviews on the platforms buyers and models both read. And make your own pages state plainly what you are best for, so the model has an easy, liftable claim rather than having to infer one. Citations get you into the room; consensus is what gets you recommended once you're in it.

This is the fragile win. The model recommends you from memory and from other people's pages, without pulling your own — so the recommendation is anchored to nothing you control and can quietly disappear when the model is retrained. The fix is supply-side: make sure the engines can actually reach and read your site, because a recommendation backed by a live, citable page is far harder to dislodge than one held in a model's head. That means crawlable pages that return real content to a fetch (many live-answer crawlers do not run JavaScript), clean structured data, and server-rendered copy that says what you do. Your server logs show exactly which AI crawlers got in and which were turned away — we walked through reading them in The Supply Side of AI Visibility.

Any tool that reduces AI visibility to a single score has already chosen for you, and it tends to choose the flattering half. Cited and named answer different questions, have different causes, and demand different work — one from your content and PR, one from your engineering. Measure both, keep them apart, sample each properly, and read the distance between them. That distance is not a vanity metric; it is the most honest to-do list the medium will give you.