ChatGPT fetched 223 pages and cited 16 on one prompt, Yeşilyurt logs
A stream capture from Metehan Yeşilyurt shows where the drop happens, and it is not at the crawl.
What Yeşilyurt captured
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 the 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.
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. He also hit rate limits by Saturday, so this is one prompt on one day.
Where the drop happens
Most coverage of the leak is reading it as a spec sheet. We think the number to sit with is the ratio. 223 pages were fetched and chunked. 16 were cited.
Being retrievable and being cited are separate problems on the same prompt. Crawlability, indexing, engine coverage and 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 stage rewards a 170-word passage that answers the query on its own, from a source the model is willing to name.
One prompt is not a pattern. We think the funnel shape is directionally right because it matches what Tomek Rudzki and others have logged at larger scale, but the exact numbers should not be quoted as a benchmark.
Before you change anything
Run your own version of the capture. Pick five prompts you care about and log, for each, the URLs cited in the answer and the passages those citations point to. Read the passages cold, without the rest of the page around them.
If a passage does not answer the prompt by itself, the fix is in the writing, not the schema. If your own pages are being fetched but not cited, you are losing at the second stage, and more crawl access will not change that.