Mention, citation, source: three things people confuse

Someone sends you a screenshot. They asked ChatGPT which tools are worth looking at in your category, and there you are, named in the second sentence, described more or less correctly. Then they open analytics, look for the traffic that screenshot should have produced, and find nothing. Not a small number. Nothing.

The usual conclusion is that AI visibility is hype. The better conclusion is that three different things have been collapsed into one word. A mention, a citation and a source behave differently, show up in different places, and demand completely different work from you. Most conversations about llm visibility never make the distinction, which is why so much of the advice in this space is unusable.

Three things, not one

A mention

The model names you inside the text of its answer. That is all. No link, no source list, no footnote. A reader sees your brand, reads a sentence about what you do, and has nowhere to click even if they want to. If they are interested they will go and search for you separately, in a session your analytics files under organic or direct, with no evidence anywhere that a model was the reason.

A mention leaves no trace in any analytics tool you own. The referral data is not messy or under-attributed. There is no referral, because there was no link.

A citation

The model names you and points at one of your pages as a source, usually as a marker in the text or a link in a source panel. Now a click is possible. Perplexity does this by design, because retrieving and attributing live sources is what the product is. ChatGPT does it when it browses, which is not always.

A citation can produce a referral, and that referral shows up in your logs with a recognisable host. People over-index on it for exactly that reason: it is the only one of the three that behaves like the web they already know.

A source

This is the one almost nobody looks at, and it is the most valuable. Sources are the pages the model is leaning on to build its answer about you. Not necessarily the pages it links. The pages it uses.

Here is the uncomfortable part: they are frequently not yours. Review sites, forum threads, comparison articles written by people who have never spoken to you, third party documentation. When a model describes your product, it is very often paraphrasing someone else's description of your product. The description you wrote, on your own site, may not be in the room at all.

Why the distinction decides what you can do

Suppose you are mentioned constantly and cited almost never. That is not a visibility problem: the models know you exist and will name you in front of buyers. What fails is something else. When the model needs a page to stand behind a claim, it does not reach for yours. Your pages are not the source it selects. Fixing that means asking whether your site answers the buyer's question in a form a machine can lift, or answers a different question in the language of a brochure.

Now the reverse: you are cited when someone asks about you by name, but never mentioned when someone asks the category question without naming you. No amount of on-site work will touch that. The model can find you when pointed at you. It does not think of you unprompted. The two failures look identical on a dashboard that counts "appearances" and have nothing in common. One is a rewrite. The other is going out and getting written about.

The part that catches people out

You can be extremely visible in AI answers and see zero clicks. Not few. Zero.

That is not a failure state. It is the normal behaviour of a mention, and mentions are the bulk of what happens. A model answering from what it absorbed in training has no link to give you, because it is not reading a page while it answers. It is generating text. No URL is involved anywhere in the process.

So a team measuring this with Google Analytics reaches the obvious conclusion, which is that nothing is happening. Something is happening. It is happening in a place that does not emit an HTTP referrer, and the instrument they are holding was built to count HTTP referrers. If you have decided AI visibility is not real because your referral traffic has not moved, you decided it with a tool structurally incapable of seeing the thing you were testing for.

What is going on under the hood

The three outcomes fall out of two ways a model can produce an answer.

From training memory. The model was exposed to enormous amounts of text, your corner of the web included, and what it retained is a statistical impression of how the world talks about your category. When it answers from this, it is recalling, not looking anything up. There is no source to cite because the model genuinely does not know where the information came from, and any link it produces in this mode is at risk of being invented. This is where mentions live. Visibility here is slow to build, slow to decay, and reflects what the web said about you before the training cutoff.

From retrieval. The model runs a search, pulls back a handful of pages, reads them, and writes an answer grounded in what it just fetched. Now there are real documents in context, and those documents can be cited. This is Perplexity's default posture and what ChatGPT does when it browses. Visibility here moves fast: a page published this week can be in the answer set next week, and a page that drops out of the retrieval set disappears just as quickly.

Retrieval shows you what is findable now. Memory shows you what stuck. Read one as if it were the other and you will draw the wrong conclusion twice.

What to do with the sources

If the pages a model draws on to describe you are mostly not yours, your work is not on your website. It is out there, on the pages that are doing the talking. That is the whole point of ai citation tracking: not a vanity count of how often a machine said your name, but a map of where your reputation is being manufactured, and by whom.

The most direct move available to you costs nothing. Look at the sources cited when a model recommends your competitors instead of you. Ask the category question, the one a buyer asks before they know any of the names, and read the source list underneath the answer. If the same review site, the same forum, the same comparison page keeps appearing behind whoever gets recommended, you have found the shortlist of places where your absence is costing you.

Doing that once by hand is useful. Doing it across a real prompt set, four platforms and every language you sell in is a job nobody has time for, which is why PSentry reports the sources and citations behind each answer alongside the mentions. Two limits are worth stating plainly, since the category rarely bothers: it maps what is already out there, twice within a month rather than by the hour, and it makes no promise about where that map will place you. Mapping a source is not the same as authoring one, so the answers themselves stay exactly as the models already give them.

The short version

A mention means they know you. A citation means they will send you someone. A source tells you who is really writing your reputation. Confusing the three is how a company concludes that nothing is happening, at exactly the moment something is.

Frequently Asked Questions

If a mention produces no click, why does it matter?

Because it lands at the moment a buyer decides what to consider at all. A model naming a few vendors in a category is doing the work a shortlist does. The click, if it comes, comes later and looks like a branded search: the influence arrives well before the traffic, which is what makes it easy to dismiss.

Can I tell whether an answer came from memory or from retrieval?

Usually, by eye. If the answer shows a source panel or inline markers pointing at real pages, it retrieved. If it is fluent prose with no attribution and it stays confident when you ask for a link, it is answering from training. Asking directly for sources is a decent test: a model working from memory will often hand you a plausible URL that does not resolve.

My competitor gets cited and I do not. Is that a content problem?

It may be a retrievability problem first. Check that your pages are fetchable, that nothing in your robots file blocks the AI crawlers, and that the answer to the buyer's question exists on your site in plain language rather than implied across three paragraphs of positioning. If all that is fine, the likelier issue is that better-placed pages elsewhere answer the question before yours gets a look.

Do the sources change between languages?

Often completely. The pages a model retrieves to answer a question in German are German-language pages, and the reputation built on them may have nothing to do with the one you carefully constructed in English. A brand that is well sourced at home can be described, in an export market, almost entirely by strangers.

Is a single check worth anything?

No. These models are probabilistic, and one answer is an anecdote. Only patterns across many prompts, repeated over time, are worth acting on, which is why a one-off screenshot, flattering or otherwise, should not be driving decisions.