Watch: What extraction method is Google using to build grounding snippets?
An analysis of Google's Gemini grounding pipeline, examining how extractive summarization selects query-focused sentences to build grounding context from web sources.
Transcript
When you ask Google's Gemini a question, how does it pull facts from the web to back up its answers? The secret lies in its grounding pipeline, specifically a process called query-focused extractive summarization.
When you enter a prompt, the system breaks your request down into single-intent queries. For each query, it pulls a handful of search results. This is where the extraction begins.
Instead of reading whole pages, the system breaks web content down into individual sentences. It scores each sentence against your query, keeps the most relevant ones, and joins them together with ellipses to build a grounding snippet. Because this process is highly query-dependent, the exact same webpage will yield completely different snippets depending on what you asked.
By examining this pipeline, we can see exactly how it behaves. First, it has a strong preference for sentences that match the query, skipping unrelated sections of a page entirely. Second, it has a heavy lead bias, meaning it almost always grabs opening paragraphs wholesale. Third, because it operates strictly at the sentence level, it sometimes sweeps up structural noise like table of contents entries, section headers, and paragraph markers.
These extraction chunks are assigned confidence scores ranging from point-one to one-point-zero. This entire behavior can be closely replicated using a fine-tuned DeBERTa model, giving us a clear window into how modern AI connects its answers to real-world sources.