222,753 searches analyzed: how Gemini searches and picks its citations
gaganghotra_ · x · 2026-10-15
RESONEO analyzed 222,753 Google searches across 76,450 public Gemini conversations. Findings: Google sends a hidden score sheet classifying questions (35 topics, 47 tasks, 22 follow-up types, 2 sentiment scores, median 2.5s); Gemini rewrites queries (55% of non-English questions add an English query, 37% add a year); Wikipedia leads citations (1 in 3 answers), news sites rose to 55/100, Reddit emerged at 5.2 while health/research sites dropped sharply; YouTube is read but rarely cited; Gemini can search without showing sources.
More from Models
- Hacker turns EmbeddingGemma 2 into 288x realtime STT and a fast TTS — kastnerkyle · 2026-10-15
- Researcher calls out Meta for releasing a model without pricing, calling it misleading — eliebakouch · 2026-10-15
- Liquid AI's LFM2.5-24B-A2B tops IFEval with just 2.45B active parameters — huggingface · 2026-10-15
- Mollick: Gemini 4 launch shows Google following trends, not its usual product innovation — emollick · 2026-10-15
- StepFun's Step 5 Preview ships open weights, demoed porting a full game codebase to TypeScript live — Aizkmusic · 2026-10-15
- Chinese Creator Says 'Opus 5.5' Skills Can Build a Channel From Scratch — xiaohu · 2026-10-15