RESONEO, an SEO research firm co-founded by Olivier de Segonzac, spent July digging through 26,900 distinct pages, close to 88,000 individual search results, and 1,200 full ChatGPT answers to trace what the assistant fetches before it answers. The firm built canary pages, decoy web pages logged at the server level to reveal exactly which crawler requested them and when, and paired that with a browser extension reading data ChatGPT sends before its interface hides most of it. This is one team’s reverse engineering of an undocumented system, not an OpenAI disclosure, and every figure below is RESONEO’s own measurement.
Until July 21, a hidden field in that stream named the pipeline behind each search result: labrador, OpenAI’s own retrieval hub covering web results plus news, academic papers on arXiv, Reddit threads, and YouTube. Two more pipelines, bright and oxylabs, feed in Google results scraped and purchased live from outside data providers. That field disappeared overnight across every account RESONEO tracks. The firm now reidentifies each pipeline through a classifier trained on formatting quirks such as snippet length and title shape, accurate about 98% of the time by its own account.
Routing between those pipelines is a cost decision, not a technical one. RESONEO found ChatGPT opens no page at all in 93% of instant-mode answers, leaning entirely on labrador’s titles and short snippets because instant mode has to answer in seconds for users who pay nothing. Thinking mode, reserved for paying subscribers willing to wait, brings in scraped Google rankings via bright and adds genuine page visits, carried out by the ChatGPT-User robot, the crawler OpenAI uses to open pages live. RESONEO notes that over nine in ten ChatGPT users sit on the free tier, which quietly makes labrador the pipeline deciding visibility in most live conversations.
The sharpest new signal concerns how much of the web a paid Thinking conversation actually reaches. Replaying identical medium-effort prompts six weeks apart, RESONEO measured search fan-outs per conversation drop by nearly half, from 3.56 to 1.90. Retrieved links fell in step, from a July average of 47.2 down to 33.9 by August, and the number of distinct domains reached slid from 21.9 down to just 15.5. Push the effort setting to its highest level and the contraction gets sharper: 67 URLs shrank to 40.5, while 29.5 domains dropped to 14.7. Fewer domains reached per conversation means the same visibility gets divided among a smaller set of sites, a concentration effect that matters more to a working SEO team than any single ranking tactic.
RESONEO’s own capture logs place that shift between July 26 and 31, before the model change OpenAI shipped on August 6. That timing rules the widely covered refresh out as the explanation. The firm does not say what actually caused the narrowing, so this is a measured before-and-after, not a diagnosed cause, and RESONEO warns its whole stack keeps shifting week to week.
Separately from the search index, ChatGPT keeps a shared cache of every fetched page, stored as converted Markdown and reused across every user and account tier no matter who triggered the fetch. A copy stays fresh for roughly half an hour. Once that window closes, users still get the stale version immediately while a background job updates it for whoever asks next. Oncrawl’s Jerome Salomon documented copies that were still being handed out well past the 90-day mark following the original fetch, with no eviction limit in sight. Setting Cache-Control to no-store on a page does not stop that caching, and a noindex tag fails to stop it too.
Two hard limits apply regardless of cache state. The fetching robot does not run JavaScript, so anything rendered client-side stays invisible to it. A page capped at exactly 4 MB is not trimmed if it runs over. The request instead comes back as a flat HTTP 400 error, and the model learns nothing about that page at all.
RESONEO’s full corpus shows how steep that funnel gets: 61,332 URLs reached the sources sidebar, only 5,032 became a citation’s lead source, and just 759 pages ever got a real open, all inside Thinking conversations. An opened page’s odds of getting cited run close to 74%. A page that was merely retrieved, never opened, converts at just 7%. Whenever the assistant reads a page unprompted inside Thinking mode, the citation shown to readers skips the usual utm_source=chatgpt.com tag entirely, so tracking referrals by that parameter alone counts the clicks and misses the reads.
Search teams should pull the ChatGPT-User agent from server logs instead of relying on UTM tagging, and treat a shrinking domain count in paid Thinking as a reason to prioritize a smaller set of pages that are demonstrably crawlable, readable without JavaScript, and comfortably under the 4 MB ceiling.
Search Engine Land published Olivier de Segonzac’s RESONEO analysis, “Inside ChatGPT’s retrieval stack,” on Aug. 17, 2026.
