How ChatGPT Chooses Its Sources: What We Actually Know

ChatGPT does not rank pages: it ranks the question first

Direct answer: ChatGPT does not assign a ranking score to web pages the way Google does. Before it even searches, it sorts your question into an internal category (instant search, shopping, local, reasoning, or plain knowledge) and that category alone decides whether a web search happens at all. A question sorted into the "general knowledge" bucket gets answered from the model's training memory, with no search and therefore no possible citation, regardless of how good your page is.

This is the part most GEO guides skip: optimising a page is pointless if the question that should lead to it never triggers a web search in the first place. A researcher who analysed ChatGPT's network traffic documented this internal sorting system (the technical parameter is called turn_use_case), and the finding fits in one sentence: wording decides the bucket, not the topic. Two questions about the same subject can take entirely different paths depending on how they are phrased. "What is an SEO audit" is likely to be treated as general knowledge; "best SEO audit tool 2026" almost always triggers an active search, because the word "best" signals a comparison to settle rather than a definition to recall.

Between the question and the answer: the real mechanics

When a web search is triggered, ChatGPT does not simply forward your question as-is to a search backend. An analysis of roughly 50,000 prompts found that the system splits the question into several sub-queries, sent in parallel to different search backends, a technique known as query fan-out. The results of those sub-queries are then fetched, filtered, and synthesised into the final answer. This synthesis step is what fundamentally separates an answer engine from a traditional search engine: you are no longer competing for a position in a list, you are competing to be the sentence the synthesis keeps.

ChatGPT's search mode relies heavily on Bing's index to fetch recent content. In practice, content absent from Bing's index starts at a structural disadvantage, regardless of editorial quality. It is a point many Swiss companies discover late: they watch Google Search Console closely, but Bing Webmaster Tools was never even set up.

The hidden pipelines behind the answer

Independent researchers (Chris Green and Suganthan Mohanadasan, work published in mid-July 2026) documented, by observing search mode's network traffic, four distinct retrieval channels operating behind the scenes, invisible in the citation cards shown to users.

One figure deserves caution rather than being treated as a law: in Suganthan Mohanadasan's sample (roughly 1,240 source records over a few days, one account), Labrador accounted for the large majority of fetches. The author himself warns that his percentages indicate a direction, not a reliable measurement: his sample is small and skewed toward SaaS and commercial queries. Another researcher, on a different sample, observed Bright playing a much larger role on commercial queries. Take away the principle, not the number: which channel gets used depends on the query type, not a universal ranking.

The same research revealed a telling gap between being fetched and being cited. In the observed sample, Reddit was fetched 278 times but cited only 11 times; YouTube was fetched 201 times and cited zero times. Being present in the context the model consults guarantees nothing: citation is an additional filter, applied after retrieval. The researchers also found that 11.6 percent of prompts changed their primary channel between repeated runs; when that switch happens, fetched-URL overlap drops by roughly 45 percent. ChatGPT's answer to the same question is not fixed over time, which is why a one-off check never tells you much on its own.

What actually determines the citation, once search is triggered

A study published by Discovered Labs, covering ChatGPT's citations, found an 87 percent alignment between ChatGPT's citations and Bing's top results: ranking well on Bing (not just Google) remains the most reliable foundation for a citation. Wikipedia alone accounts for nearly half (47.9 percent) of citations among ChatGPT's ten most-cited sources, confirming a clear preference for encyclopedic-style content with clear definitions and well-identified entities.

A second figure qualifies the first rather than contradicting it: an Ahrefs analysis of 15,000 queries measured that only 12 percent of URLs cited by AI tools also appear in Google's top 10 for those same queries. Both figures coexist because they measure different things: alignment with Bing is about the index used for retrieval, overlap with Google is about the final ranking. A page can be indexed and retrievable without ever topping a classic Google ranking, and still end up cited.

Content freshness also plays a measured role, documented this time on Perplexity: content updated within the last 30 days receives 3.2 times more citations than older content on the same topic. Nothing guarantees an identical ratio on ChatGPT, but the underlying principle (dated, regularly revised content earns more trust from a synthesis system) shows up across most answer engines. A clearly identified entity (company name, location, offer) tagged with structured data also makes it easier for the model to decide whether a page precisely answers the generated sub-query, a topic we cover separately in our article on the Schema.org types that actually matter for answer engines.

Fetched, cited, mentioned: three different outcomes

A page can face three distinct fates with ChatGPT: being fetched into context without ever surfacing to the user, being cited as the source behind a specific sentence with a visible link, or having the brand mentioned with none of its pages serving as a source at all. These are three very different levels of outcome, and conflating them leads to over- or under-estimating your actual visibility in answer engines.

The first case (fetched, not cited) is the most common and the least visible: it is the fate of most of the Reddit and YouTube pages mentioned above. The second case (cited with a link) is what GEO tracking tools actually measure. The third case (brand mentioned without a source) comes from the model's training memory: your name can show up in an answer built entirely from what the model learned during training, with no web query involved at all, a track that is completely independent of your current pages.

When this should not be your priority

If most of your potential traffic matches questions sorted internally as "general knowledge" (definitions, history, stable factual questions), optimising your pages for citation will not change anything: those questions never trigger a web search, the answer comes entirely from training memory. In that case, effort is better spent on classic Google SEO, which indirectly shapes what the model learned during its last training run, rather than on real-time citation optimisation that will never even be consulted.

Likewise, if your business runs on very local, transactional queries (a tradesperson, a local shop with no informational content), the Oxylabs channel and business listings matter more than blog article structure. Publishing long-form content purely to chase a citation then costs more than it returns: it is better to consolidate your presence on professional directories and reviews, which feed that channel directly.

What you can actually do

1. Get well indexed on both Google and Bing. The two indexes feed different pipelines; neglecting Bing cuts you off from a channel aligned with 87 percent of observed citations.
2. Write standalone passages. A definition or answer that keeps its meaning out of context is more likely to survive into a synthesis than a paragraph that assumes the reader has read the whole article.
3. Make content accessible without client-side JavaScript execution. A bot that cannot fetch the content can neither fetch nor cite it (see our article on the server-side rendering trap).
4. Build a presence on third-party sources with strong editorial authority (trade press, recognised professional directories) rather than relying solely on your own domain's authority.
5. Refresh your highest-value content regularly, with a visible date: it is a measurable freshness signal, not just an editorial habit.

What this means for a Swiss SME

The practical conclusion is not that classic SEO should be dropped in favour of GEO, or the reverse: the two pipelines largely feed off the same indexes. A page's Google and Bing ranking remains the best predictor of its odds of eventually being fetched and cited. The difference then plays out in structure: a well-ranked page written like a marketing brochure will get fetched, then discarded at synthesis time, for lack of a citable sentence. For a full overview of the six levers shaping this visibility, our overview of GEO and AI answer engines remains the starting point, and our analysis of GPTBot, ClaudeBot and PerplexityBot completes this mechanism on the technical access side. For a Swiss SME that wants to know where it stands before spending time on this, the GEO Sprint starts from a concrete audit rather than a guess.