How does ChatGPT pick its sources?

Ask ChatGPT for a recommendation and it might answer instantly from what it already knows, or it might search the web first and cite real pages. Only one of those modes involves picking sources at all — and understanding which is which explains why the same question can get different answers.

Two ChatGPTs

There's the version that answers from its training data: no search, no citations, just a synthesis of what the model learned before its cutoff, delivered with total confidence and no way to check where it came from. And there's search-grounded ChatGPT, which runs a web search, reads a set of results, and builds its answer from what it finds — with links to the actual pages it used. When people talk about being "cited by ChatGPT", they mean the grounded mode. That's the one worth optimising for, because it's the only one where being a good source can actually change the answer.

What it looks for

Grounded ChatGPT behaves a lot like a careful researcher skimming search results rather than reading every page in full. It favours pages that directly answer the question asked, rather than pages that merely mention the topic in passing among a dozen other subjects. It looks for entities it can recognise and describe consistently — a business whose name, description and details read the same wherever it turns up, rather than a slightly different pitch on every page. And it weighs corroboration: a claim repeated across several independent sources gets more confidence than the same claim made once, on a single page, by the business itself.

Why answers vary

Ask the same question twice and you can get different brands named, even minutes apart. Each search pulls a fresh set of results, and the model samples from what it finds rather than running a fixed lookup — so no single answer is the whole picture, and treating one good response as proof you're "winning" is a mistake. What matters is the trend across many runs, not any one response. That's why tracking visibility means asking repeatedly and watching the pattern, not screenshotting one lucky answer and moving on.

Which pages win citations

Homepages rarely get cited, because they rarely answer one specific question well — they're built to introduce a business, not settle a query. A specific, useful page — one built to directly answer a real question, with the answer near the top — is what the model reaches for instead. If your site's best answer to "how much does X cost" lives three clicks deep with no clear heading, that's the fix, not blanket authority-building on the homepage. Checking which of your pages are actually earning citations tells you whether your content is already shaped the way the model wants it, or whether the useful information is buried where it can't be found.

Track it rather than guess

Manually running the same prompts over and over, across sessions, isn't something you can keep up by hand — and a handful of spot checks, run once and not repeated, won't tell you much on their own about where you actually stand. Tracking ChatGPT specifically gives you the repeated sampling this actually requires, and turns scattered impressions into a pattern you can act on.