An AI assistant can name a page that does not rank for the question, but the case where that happens is narrow: the assistant answered from its own weights instead of running a search. When an assistant does run a search, the page has to be findable in whatever index that assistant reads before it can be quoted, which is why ranking is a prerequisite for citation in every retrieval-backed answer. The two mechanisms produce different artefacts, and the difference is visible from the outside. Retrieval produces a link. Memory produces a name with nothing behind it.
Before an answer engine can quote a page, three things have to be true, in order. The page has to be in the index the engine reads, and that index is not always Google's — for ChatGPT search it is Bing's index. The page has to come back for the query the engine actually issued, which is frequently a rewritten version of what the person typed rather than the words on screen. And the passage the engine wants has to be in the HTML that was fetched, not assembled by a script after load. A page that fails the first of those cannot be quoted no matter how well the third is handled.
A retrieved citation and a remembered mention look similar in an AI answer and behave nothing alike. A retrieved citation came from a document the engine fetched during the answer, so it carries a URL that resolves and it can contain facts newer than the model. A remembered mention came out of the weights, so it carries a name and often no URL, it cannot contain anything the model was not trained on, and asking for the source produces a plausible address that may return 404. Five signals separate them.
| Signal | Retrieved during the answer | Recalled from training |
|---|---|---|
| A link beside the claim | Present, and it resolves | Absent, or offered and broken |
| Facts newer than the model | Can appear | Cannot appear |
| Asking where it came from | Names a document it fetched | Names a plausible address |
| Re-asking in a fresh session | Sources move as the index moves | The same wording tends to return |
| Your server logs | May show a fetch from a named crawler | Show nothing at all |
None of the five signals is conclusive on its own, and the first one is the one worth checking first. An answer that names a company and offers no link has not read that company's site during the answer, whatever else it did.
No setting anywhere grants a citation. What a small site controls is the shape of its passages and its own server logs. What makes a paragraph cheap to quote, how to tell whether assistants are fetching your pages, and why a cited page can still show zero referrals.
Market4 turns one release note into a changelog page, a blog post, a mail-out and a week of social posts — and then tells you which of them brought anyone back.
A mention with no link produces no referral, no landing page and no row anywhere in your analytics, because no request was ever made to your server. This is the harder half of the zero-click problem: a zero-click impression at least leaves an impression row in your own Search Console, whereas a name spoken inside a model's answer leaves no artefact on your side at all. If a founder tells you an assistant recommends their product, the honest position is that this is a report of one session, not a measurement, until it is reproduced.
Wanting to be quoted by an answer engine is a retrieval problem before it is a writing problem, and the order matters because effort spent on the second while the first is failing returns nothing. The work divides cleanly:
Being quoted without ranking is not a strategy anyone can execute, because the mechanism that produces it is the model's training run and nobody outside the lab schedules that. The mechanism you can act on is retrieval, and retrieval is the one that also sends you a visitor.
Yes, in two situations. The first is that the engine reads a different index — ChatGPT search reads Bing, so a page that ranks in Bing and not in Google can be retrieved and cited. The second is that the answer was produced without a search at all, from the model's training data, in which case there is usually no link and nothing to click. Neither situation is a substitute for ranking in the index the engine actually queries.
Check whether the claim carries a link that resolves, and whether the answer contains anything that happened after the model's training cut-off. A fetched answer can cite a page published last week; a remembered one cannot. Asking the assistant where the information came from is also informative: a fetched answer names a document it actually retrieved, while a remembered one tends to produce a plausible URL that may not exist.
No. Search Console reports impressions and clicks from Google Search, so an assistant naming your product in prose creates nothing there. If the assistant did fetch your page during the answer, the request may appear in your server logs under a named crawler user agent, and a click on a citation link will appear in your analytics as a referral. A mention with no link produces neither.
The work overlaps almost entirely, so it is not a separate budget. Being in the index, answering the question in the first hundred words, keeping the answer in server-rendered HTML and writing self-contained passages all help a page rank and all help it be retrieved. What does not transfer is chasing a mention with no link, because it cannot be verified, cannot be repeated on demand and sends nobody to the site.
The bandwidth you save by blocking an AI crawler shows up in a log. The citations you lose do not show up anywhere, because nothing reports an answer that did not mention you.
There is no markup that gets a page into an AI Overview. Eligibility is ordinary indexing plus permission to show a snippet, which makes the question what to write rather than what to add.
