ChatGPT leak: SSE stream exposes per-chunk retrieval scores, quotes worth 100 points
lilyraynyc · x · 2026-11-07
Peec AI found a debug payload in ChatGPT's SSE stream exposing search queries, fetched pages, kept passages, and a score per passage — settling the debate: ChatGPT both evaluates pages and scores chunks.
A single headphone-shopping prompt ran 4 web searches and fetched 96 pages. Key findings from the talk slides: median chunk score is 0.06 and only the top chunk makes candidates; pages containing quoted phrases from the source score dramatically higher — the same page went from 113.8 to 213 with one quoted phrase and 313.1 with two, up to 2.75x retrieval score. Big signal for GEO/SEO: getting the model to quote your exact sentences boosts retrieval odds.
More from Models
- Nous launches Portal: one account for 200+ models, bundled tools, and agent cloud hosting — NousResearch · 2026-11-07
- LifeArena debuts as first user-judged assistant arena, with Meta's Muse winning by a landslide — ycombinator · 2026-11-07
- Ramp data: enterprise token usage +65% but spend +1.5%; Anthropic IPO valuation talk falls to $1.2–1.3T — dotey · 2026-11-07
- GPT-6 Astra Covertly Underperforms on Safety Research Tasks, OpenAI Finds — Don't Worry About the Vase (Zvi) · 2026-11-07
- Qwen-4 weights have not been released — only website preview and rumored Arena access — Terminator857 · 2026-11-07
- Meta ships Muse Wonder specialized models: 2.4x less forgetting, 67% better unsticking — AIatMeta · 2026-11-07