How ChatGPT decides which sources to quote
An assistant that browses does three things: it rewrites your question into searches, it fetches a handful of results, and it reads them for a passage that settles the question. A page has to survive all three steps, and most pages fail at the third.
Nobody outside the providers has the ranking function, and any article that gives you one is inventing it. What can be observed, by asking a great many questions and looking carefully at what came back, is the shape of the pipeline — and the shape is enough to act on.
Step one: your question becomes several searches
A person asks something long and conversational. The assistant does not send that to a search index. It decomposes it into two or three ordinary-looking queries, often in the language of the question, sometimes in English as well, and runs them.
The practical consequence is that classical search ranking still gates everything. If none of those searches surfaces your page in the first handful of results, the assistant never sees it, and nothing about how beautifully your answer is written matters.
Step two: a few pages get fetched
Not fifty. A small number, usually somewhere between three and ten, fetched by a crawler you can see in your logs. Several things can make a fetch useless at this point:
- The content is rendered by JavaScript after load, so the fetched HTML is an empty shell.
- A consent banner or an interstitial is the only thing in the returned markup.
- The page is slow enough that the fetch is abandoned.
- The robots file disallows the assistant's crawler, or its training crawler, depending on what that provider honours for which purpose.
- The page returns a soft 404 or redirects to a homepage.
None of these are exotic. Between them they account for most of the cases where a company is genuinely the best source on a subject and is never quoted on it.
Step three: the reading, which is where most pages lose
Now the model has three to ten documents and a question. It is looking for a passage that answers the question outright. What tends to get selected has a recognisable character.
| Gets selected | Gets skipped |
|---|---|
| Answers in the first paragraph | Warms up for four paragraphs |
| States a condition: "if X, then Y; otherwise Z" | Says "it depends on your situation" |
| Names figures, dates, versions, jurisdictions | Speaks in general terms throughout |
| Sections that stand alone | Sections that refer back to "as mentioned above" |
| First-hand: we do this, this is what we see | Aggregated: here are twelve tools |
| Shows when it was last revised | Undated, on a subject that moves |
Assistants do not reward the best article. They reward the most liftable paragraph.
What about the citations shown under the answer?
Treat them as a partial view. An assistant may read a page, use what it said, and cite a different source for the same statement — or cite nothing, because the statement is one it considers general knowledge. The absence of your link under an answer does not mean your page was not read, and its presence does not mean it was the decisive source. This is the single most frustrating property of the medium and there is no way around it.
What you can actually check
Assistant crawlers identify themselves. You can see which of your pages are being fetched, how often, and whether the fetches are succeeding. This is the most under-used source of information about AI visibility that most companies already have.
Request your own page the way a crawler would — no JavaScript execution — and read what comes back. If the answer is not in there, it is not in there for anybody.
Ask the assistants your buyers use, monthly, with buyer phrasings, in the right languages. Record what is said and which sources are shown.
Questions this settles
Assistants have used various search backends over time, and the arrangement has changed more than once. The durable point is that a search step happens against a conventional index, so conventional ranking decides what is available to read.
It depends which crawler does which job for that provider, and providers document this separately for training and for live retrieval. Blocking the training crawler and allowing the retrieval one is a common, deliberate position.
Usually because the thinner page answers the question in its first paragraph and yours answers it in the sixth. Retrieval reads for the passage, not the document.