DEJAN: Google's grounding - chunk sizes from Vertex AI Search
- th3s3rp4nt
- 20. Dez. 2025
- 1 Min. Lesezeit
Key Takeaways:
Dan Petrovic analyzed 7,060 queries via Google's Vertex AI Search that likely powers Gemini
Chunk size: The total words retrieved for grounding is surprisingly stable throughout queries at a median of 2.000 words (within 1.500 to 3.000 words)
This budget is devided among sources based on their relevance ranking:
First position with 28% to 5th position with 13% share.
77% of pages get 200-600 words selected. The typical page gets ~377 words.
Longer, more extended pages do not get a higher share or chunk size but rather a lower coverage (less % of their content included)
Conclusion:
Optimize your content in "chunks" with a high information-depth and highly concise structure - clear, on-point passages without filler sentences - every word should add meaning, every sentence should be easy to cite
To optimize coverage do not build extensive pages but rather ones limited to a certain topic in detail with up to 800 words only
Percentile | Total Words (Grounding) Per Query |
p25 | 1,546 |
p50 (median) | 1,929 |
p75 | 2,325 |
p95 | 2,798 |
This [grounding] budget is remarkably consistent regardless of how many sources are used or how long the individual pages are.
Longer pages do not add to longer grounding chunks
Page Chars | Avg Grounding Chars | Coverage |
<5K | 2,127 | 66% |
5-10K | 3,024 | 42% |
10-20K | 3,363 | 25% |
20K+ | 3,574 | 12% |
Sources:
