When a model cites a page, the instinct is to read it like a search result: we ranked, therefore we won. That reading gets you into trouble quickly, because a citation is answering a different question from the one a ranking answers.
A ranking is a claim about relevance across every query like this one. A citation is much smaller: it says that for this question, at this moment, this page was reachable and useful enough to lean on.
Ask the same question an hour later, or phrase it slightly differently, and the receipt can name a different page. That is not a bug in the model or a failure of your page. It is what happens when the answer is assembled fresh each time rather than looked up from an index.
Something worth sitting with: the cited page is frequently not your product page. It is a support article, a comparison you wrote two years ago, a paragraph buried in your documentation that happens to state a fact plainly.
Models reach for text that answers the question directly. Marketing pages are built to persuade, and persuasion reads as hedging when what is wanted is a fact. The pages that get cited tend to be the ones that say a thing and stop.
You can be cited in an answer that recommends a competitor. Your comparison page is a fine source for what your competitor costs, and a model will happily use it that way. The link is yours; the recommendation is not.
This is why citation counts on their own are a thin measure. A page cited as supporting evidence for someone else's pitch is not the same as a page cited because you are the answer, and a number that treats them identically will tell you a comforting story.
Three things, reliably.
It tells you the model could reach the page — no login wall, no rendering problem, nothing blocking the crawler that fetches it. That sounds trivial until you find the page that should be cited everywhere and never is.
It tells you the page answered something. Not necessarily the whole question, but a piece of it, in language that could be lifted.
And it tells you where you sit in the shape of the answer. A citation in the opening sentence and a citation in a footnote-ish aside are both citations, and they are not remotely the same outcome.
Look at the sentence the citation is attached to, not just the fact of the link. Read what the answer is claiming at that moment. That sentence is the model's summary of what your page is for, and it is frequently a more honest description of how you come across than anything on your homepage.
If the sentence is wrong, that is the thing to fix — and it is usually fixed on the page, not in the prompt.
One citation is an anecdote. The same page cited across many different questions is a fact about how your site is understood.
If one support article is carrying a disproportionate share of your citations, that article has become your de facto product page as far as models are concerned. It is worth reading it as a stranger would, because it is doing more positioning work than anything you commissioned.
The inverse case is more common and easier to miss. A page you consider central that is never cited anywhere is not being rejected on quality — it is usually failing on something duller. It may be slow, it may hide its facts below a lot of persuasion, or it may simply have nothing quotable in it.
Where the citation sits changes what it is worth, and this is the dimension most citation counts throw away.
A source used to support the opening definition is shaping the frame of the whole answer. A source appended to a caveat near the end is decorating a conclusion that has already been reached. Both increment the same counter, and only one of them meaningfully affects what the reader takes away.
When you review answers by hand — which is still worth doing periodically, however good your tooling — read for position before you read for count.
If a model cites your page for a claim your page does not make, the instinct is to treat it as a model error. Occasionally it is. More often the page is genuinely ambiguous in a way that is invisible to anyone who already knows the answer.
Read the sentence the model produced, then find the passage it came from, and ask whether that passage could reasonably be read that way by someone with no context. The answer is usually yes, and the fix is usually a sentence rather than a rewrite.
Contents
Not your homepage. A working theory of the sources that shape what AI answers say about a brand, and what that means for where you spend effort.
Read →What the lines mean, which movements are real, and the three ways this chart gets misread.
Read →The channel where your brand is decided but no traffic arrives, and why the usual analytics can't see it.
Read →Tell us your domain and we'll send you a visibility report — free, no card.