Google antitrust trial filings reveal search pipeline: deep models only rerank top 20-30 docs
GoSailGlobal · x · 2026-11-01
Compiling evidence from the DOJ v. Google antitrust trial and leaked internal API docs, the author maps Google's full search ranking pipeline: inverted-index intersection pulls tens of thousands of docs, coarse ranking with topicality/PageRank/Navboost narrows to a few hundred, RankBrain and DeepRank only tune the top 20-30 docs (too expensive to run on more), twiddlers (PXR0357) rerank results, Tangram assembles the page, and clicks plus rater scores flow back into training. The post also compares how openly Bing, Baidu, Yandex, Naver and SEO tool vendors document each stage, and frames GEO as the equivalent chain for AI answers.
More from Infra
- Intel EMIB packaging yield at 60-65% vs TSMC CoWoS's 98%, report says — rickasaurus · 2026-11-01
- SpaceX guides to $35B quarter as Grok Bot hits 800K paid weekly users — elonmusk · 2026-11-01
- "IaC is dead": AI will manage infra changes, and Datadog-style UIs may suffer — thedealdirector · 2026-11-01
- Switching to EPYC: 4x Faster Local LLM Decode with Full Tuning Notes — mrgreatheart · 2026-11-01
- Korea lost $1.1 trillion in market cap: how a memory-chip duopoly made KOSPI an AI-leveraged node — ChinaTalk · 2026-11-01
- Nvidia engineer on Open Agent Safety Platform: verifiable, monitorable agentic AI — LuizaJarovsky · 2026-11-01