Reverse-engineering how Google builds Gemini grounding snippets: extractive, lead-biased, noisy
dejanseo · x · 2026-11-06
SEO researcher Dan Petrovic reverse-engineered Google's Gemini grounding pipeline by inspecting raw groundingSupports and groundingChunks from the API, focusing on how snippets are extracted from retrieved pages.
Key findings:
- Query-focused selection: sentences semantically close to the query are strongly preferred; unrelated sections on the same page are skipped entirely.
- Heavy positional/lead bias: opening paragraphs are extracted almost wholesale regardless of relevance.
- Structural noise ingestion: ToC entries, section headers, link-code artifacts and ¶ markers are treated as sentences and scored alongside prose.
- Sentence-level granularity, not passages or paragraphs.
- Per-chunk confidence scores observed in the 0.1–1.0 range.
Petrovic also fine-tuned microsoft/deberta-v3-large, which reproduces Google's behavior fairly closely, and published a demo, a full pipeline diagram, and annotated examples showing exactly which sentences get extracted versus skipped.
More from Models
- User screenshot suggests Anthropic's weekly Claude limits feel effectively unlimited — StewartalsopIII · 2026-11-06
- Leonie Shares Full Slide Deck on Building Decision Models, Still Early Days — helloiamleonie · 2026-11-06
- Liquid AI open-sources d1 decision models: a 2B model purpose-built for classification calls — helloiamleonie · 2026-11-06
- No RL needed: temperature scaling is the simple trick teams use to calibrate model confidence — helloiamleonie · 2026-11-06
- Freshly trained decision models are overconfident: claims 90% sure but only 68% correct — helloiamleonie · 2026-11-06
- Liquid AI lifts the decoder's causal mask: bidirectional attention beats causal for decision models — helloiamleonie · 2026-11-06