Grounding snippet · Extract sent to the model
A grounding snippet is the set of sentences an AI engine extracts word for word from a page to rely on when writing its answer. The model does not read the whole page: it receives these few sentences, often stitched together with ellipses, and most of the text never reaches it.
Standalone definition · 51 words
How does Google build a grounding snippet?
According to Dejan’s analyses, extraction is extractive, not paraphrased: each sentence is scored against the query, the best ones are kept and stitched together with “…”. Opening paragraphs are almost always kept, and the same page yields different extracts for each query fan-out sub-query.
How big is a grounding snippet in Google?
Across 7,060 queries analysed by Dejan (December 2025), Google allocates a budget of about 2,000 words per query, split between sources by rank. Source #1 gets about 531 words, #5 about 266. A typical page gets about 377 words selected, with a plateau around 540 words.
How should you write to end up in a grounding snippet?
Put the answer in the first sentences, write sentences that stand alone and avoid long intros. A dense 800-word page gets more than half of its text selected, against about 13% for a 4,000-word page in the same study.
Frequently asked questions about Grounding snippet
Is a grounding snippet the same as a featured snippet?
No. A featured snippet is shown to the user in the SERP. A grounding snippet is invisible: it is sent to the model as raw material.
Do other AI engines work the same way?
Dejan sees the same ellipsis-stitched sentence format in Claude. Details vary by engine and are not officially documented.
Terms related to Grounding snippet
Further reading on Grounding snippet
- Is chunking a technique to optimize for SEO and GEO?
- Techniques to optimise for AI Overviews and LLMs
- ← Back to the GEO glossary (126 terms)
Sources for the Grounding snippet definition
- Dejan, “Grounding Snippets” (blog)
- Dejan, “How big are Google’s grounding chunks?” (blog, 24/12/2025)