ChatGPT fetched 223 pages and cited 16, per Yeşilyurt stream capture
A weekend of pulling ChatGPT's server-sent events shows the retrieval funnel in numbers.
What happened
Metehan Yeşilyurt posted on LinkedIn on 7 September that he pulled a debug view out of ChatGPT's server-sent events over a weekend of testing. For one prompt, "best AI visibility tools", the stream logged 5 search rounds, 18 rewritten queries, 50 engine calls, 228 results, 223 URLs fetched with chunks selected, and 16 citations in the final answer.
The capture is a single prompt on a single day, and Yeşilyurt says he hit rate limits by Saturday. He describes roughly ten internal engines, per-result and per-chunk scoring, a should_fetch decision, and a second ranking pass inside the model's reasoning.
What we think is going on
Most coverage of this leak is treating it as a spec sheet for the pipeline. We think the number that matters is the ratio. 223 pages were fetched and chunked. 16 were cited.
Being retrievable and being cited are two different problems on the same prompt. The industry is mostly working on the first one. Crawlability, indexing, engine coverage, schema, all sit before the 223.
The drop from 223 to 16 happens after that work is done. Yeşilyurt's stream shows chunks of roughly 170 words being scored, with one to three kept per source. A second pass then re-orders domains inside the model's reasoning before the answer is written.
That second stage rewards a different thing. It rewards a 170-word passage that answers the query by itself, from a source the model is willing to name. Schema and crawl budget do not help a weak passage survive it.
One prompt is not a pattern. We think the ratio is directionally right because the funnel shape matches what Tomek Rudzki and others have logged at larger scale, but the exact numbers should not be quoted as a benchmark yet.
What to do about it
Run your own version of the capture before you change anything. Pick five prompts you care about, log the cited URLs, and count how many distinct domains actually appear across the set. That is your own fetched-to-cited gap, and it is the number to move.
For pages you already know are being retrieved but not cited, read them in 170-word windows. Ask whether any single window answers the prompt without the rest of the page around it. Rewrite the ones that do not, for the reader, not for a chunker.
Check what your meta description and title actually say, because Yeşilyurt found the raw meta_tags list travels with every result while og_data was empty. That is one input into the per-result score, alongside the body render.
Keep the crawl and schema work going. It is what gets you into the 223. Do not treat it as the finish line.
What we're watching
Yeşilyurt has said more posts are coming this week from him and others working the same stream. What would firm this up is a second capture, on a different prompt class, showing a similar fetched-to-cited ratio and the same two-stage ranking. What would break the read is evidence that the 16 citations were drawn mostly from a small pre-ranked shortlist, with the other 207 fetches used only for grounding checks.