Emulent Let’s Talk
Skip links
Emulent Diagram

How ChatGPT Chooses What Sources to Include

ChatGPT is a language model first: by default it answers from training data with no sources at all. Citations only exist when its web search tool fires. When it does, the pipeline looks like this, and most retrieved pages never make it into the answer.

01

Decide whether to search the web

Everything hinges on whether the search tool is invoked. No search, no real sources.

Default: answer from memory

Most conversations are answered from the model’s training weights, with no retrieval and no citations. If it appears to reference a source in this mode, it is reconstructing a plausible reference from memory, which is where fabricated citations come from.

Triggers for live search

Search fires when the model judges the answer needs current or verifiable information, or when the user forces it. Time-sensitive topics and commercial or research intent (words like best, reviews, comparison, pricing, or a year) trigger it far more often than general questions. Only a minority of conversations search at all.

02

Write its own search queries

The user’s prompt is not what gets searched. ChatGPT rewrites and decomposes it.

Query fan-out

A single prompt is split into multiple sub-queries, each fetching its own set of candidate URLs. One question can pull from dozens of different search results, so pages covering adjacent subtopics get more entry points into the pool.

The index behind it

ChatGPT’s search runs on Bing-powered search infrastructure plus OpenAI’s own crawling (OAI-SearchBot). A page not indexed in Bing is effectively invisible to it, whatever its Google rankings say.

03

Retrieve and read candidate pages

Candidate pages are fetched and their content extracted. What survives this stage depends on access and authority.

Open pages only

It reads open-access pages and skips paywalled or blocked content. Sites that block OpenAI’s crawlers in robots.txt take themselves out of the pool.

Authority as a gate

Analyses of large query samples find domain authority, content quality, and platform trust dominate which pages get pulled. ChatGPT is risk-averse: heavily referenced domains are several times more likely to be retrieved, a pattern researchers call the trust cliff.

Ranking helps, loosely

Ranking well in classic search correlates with retrieval but does not decide it. A large share of AI citations come from pages outside the top 10, because fan-out sub-queries reward narrow, specific coverage.

04

Cut most of them: retrieval is not citation

The steepest filter in the pipeline. Studies of large prompt samples find roughly 85 percent of retrieved pages are never cited in the final answer.

Claim support wins

A page gets cited when it directly supports a specific claim in the answer being written. Clean, self-contained statements of fact, comparison tables, and extractable lists beat pages where the point is spread across paragraphs.

Evidence density

Pages carrying concrete evidence get cited more: one large study found pages with expert quotes and rich statistical data averaged roughly double the citations of pages without them.

A handful survive

A typical grounded answer carries about three to six clickable citations. Being mentioned by name and being cited with a link are different outcomes: mentions can come from training data, citations only from this pipeline.

05

Compose the cited answer

The surviving sources are synthesized into one response, with inline attribution.

Inline numbered citations

Claims drawn from the web carry small numbered or favicon references that link to the source page, placed next to the statements they support.

Sources panel

A sources list collects every cited page for the response, so users can audit where each part of the answer came from and click through to the originals.

Fresh every time

The same question asked twice can search differently, retrieve differently, and cite different pages. Nothing about the source set is fixed; it is re-decided per response.

The contrast with Gemini: ChatGPT’s candidate pool comes from Bing’s index and OpenAI’s crawler, while Gemini’s comes from Google Search and the Knowledge Graph. Companion diagrams: how Gemini chooses sources and how Google ranks pages.

Want your pages to be the ones that get found and cited?We build SEO, AI search optimization and content strategy into every site we design.

Get a Free Quote