Is ChatGPT using your work without giving you credit?

Your website could appear in ChatGPT’s searches, have passages extracted and scored, and still receive no citation in the answer.

That is the uncomfortable possibility illustrated by findings from Peec AI researcher Metehan Yeşilyurt. In one example, he says, ChatGPT fetched content from 223 URLs while answering a question about the best AI visibility tools. The finished answer contained 16 citations.

For a publisher, that means an absent link cannot tell you whether ChatGPT missed your work or found it and moved on.

Yeşilyurt says he discovered detailed retrieval records in the stream of data ChatGPT sends to a browser while generating an answer. In his LinkedIn post, he describes five search rounds, 18 different queries and 50 calls to search systems behind that single response.

The records allegedly go further, showing scores for individual results and passages, decisions about which pages to fetch, and which extracts were selected for the model.

Search Engine Watch reviewed screenshots of the post and accompanying examples. It has not independently reproduced the capture or examined the full dataset. The material does not establish that every fetched page influenced the answer, or that every ChatGPT search follows this pattern.

The page extracts raise another concern: how much of your work survives the process?

According to Yeşilyurt, retrieved pages are converted into text resembling Markdown and divided into passages of roughly 170 words. In the behavior he describes, the model receives one to three selected passages per source.

One displayed example reduces a product page to headings, copy, labels for an email field and a button, and an image placeholder. The visual layout is gone.

Consider what that could mean for a detailed product test. The result may sit in one passage, the test conditions in another, and the limitation that changes the recommendation further down the page. If only part is selected, the context that made the original useful could disappear. The screenshots do not demonstrate that this happened in the final answer, but the reported extraction process makes it a question worth testing.

There is no demonstrated fix in the findings. The roughly 170-word figure is an observation attributed to the researcher, not evidence that publishers should rewrite every section to that length. Metadata appears in the displayed records too, but its presence does not establish its influence on citation selection.

The broader gap between retrieval and citation is already acknowledged in OpenAI’s web-search API documentation. It says the list of consulted sources is often longer than the inline citations. That supports the distinction, though it does not verify this ChatGPT capture.

The research also described search-source labels and specialized indexes observed in ChatGPT’s server events. Yeşilyurt’s latest claims add a more detailed picture of the selection happening between finding a page and presenting an answer.

For publishers, the troubling part is how little the final answer explains. Your page may have been discovered, retrieved and considered. Without a citation, the reader has no direct route back to it from that answer.

And the publisher is left trying to fix a disappearance without knowing where it happened.