How to Win Google's AI Slice for eCommerce

[How big are Google’s grounding chunks? by Dan Petrovic in AI SEO](https://dejan.ai/blog/how-big-are-googles-grounding-chunks/) Now let’s translate what it means for brands trying to win in agentic eCommerce, where the “shop shelf” is increasingly a chat answer, a comparison table, or an AI-generated shortlist. The uncomfortable truth: AI search is not “reading your page” Traditional SEO trained all of us to think in total page quality. Word count, comprehensiveness, topical authority, internal links, the whole thing. But Gemini-style answers are built from grounding snippets. That means Google is pulling small, extractive chunks of text from a handful of sources, then using those as the evidence for the final answer. Petrovic’s analysis is useful because it quantifies that behaviour at scale: 7,060 queries analysed 883,262 grounding snippets Average chunk length: 15.5 words So yes, you can publish a 4,000-word masterpiece. But the model might only “see” a few hundred words of it, and often not the bits you laboured over. The headline number: Google appears to run on a ~2,000 word grounding budget Across those queries, Petrovic found a remarkably consistent “budget” of grounding text per query: Median total grounding words per query: 1,929 Typical range: roughly 1,546 (p25) to 2,325 (p75) 95th percentile: 2,798 In plain terms: there is a ceiling. The pie does not get bigger just because your page is longer, or because there are more sources. This matters for eCommerce because a lot of brand teams are currently doing the same thing they did for classic SEO: publishing longer guides, adding more FAQ blocks, stuffing “helpful content” everywhere. If the model is budgeted, the game becomes: how do you earn more of the budget? Ranking decides your share of the pie Petrovic’s data shows the budget is split across sources by relevance rank. The higher you rank in the grounded sources list, the more words you get included. Median words by rank looked like this: #1 source: 531 words (28%) #2 source: 433 words (23%) #3 source: 378 words (20%) #4 source: 330 words (17%) #5 source: 266 words (13%) This is the bit most brands miss. You are not competing to be “one of the sources”. You are competing to be one of the top sources, because the difference between #1 and #5 is roughly double the grounding real estate. For agentic commerce, where an assistant might only cite a few sources, this becomes existential. If you are source #5, you might be present but effectively invisible. The per-page reality: you typically get ~377 words When Petrovic looked at individual pages, the typical amount of text selected from a single source was around: Median: 377 words Most pages: 200 to 600 words A practical plateau around: ~540 words / ~3,500 characters That 377-word figure is one of the most important numbers eCommerce teams can internalise. Because it forces a hard question: If Google only “reads” ~377 words of my page, which 377 words do I want it to read? - If you do not design for that, you are leaving it to chance. - Why long eCommerce pages can underperform in AI answers - Petrovic also measured how coverage drops as page size increases: Pages under 1,000 words got far higher coverage Pages over 3,000 words saw very low coverage percentages Example takeaway from the study: a tight 800-word page can get 50%+ coverage, while a 4,000-word page might get around 13% Dejan AI **This does not mean “short content always wins”.** It means density beats length when the system is extractive and budgeted. For eCommerce, a lot of pages are bloated in exactly the wrong way: - repetitive brand fluff - generic “why choose us” - long intros before the actual product facts - FAQs that never answer the buyer’s real objections - paragraphs that say a lot but prove nothing AI systems love the opposite: explicit answers, grounded facts, tight language, scannable structure. **What this changes for agentic eCommerce** Agentic eCommerce is when the buyer outsources the work to an assistant: “Find the best protein powder for sensitive stomachs under £30” “Compare these three air fryers for a family of four” “Which moisturiser is closest to CeraVe but cruelty-free?” In those flows, the assistant needs: - clear constraints - specs - evidence - comparisons - caveats (who it is for, who it is not for) - pricing and availability context And it needs them in chunks that can be safely quoted. So your job becomes less “write a beautiful page” and more “publish a set of quotable truths”. A practical playbook: design the “grounding layer” of your pages Here’s a simple way to build pages that win grounding selection without turning your site into robotic sludge. **1) Put the answer first, not the story** If you sell a product, the grounding layer should include: - what it is - who it is for - key differentiators - core specs - proof points (certifications, materials, compatibility, testing standards) - primary objections, answered cleanly Do this high up the page, before the fluff. **2) Write in quotable units** Remember the average chunk is about 15.5 words. That screams one thing: clean sentences. Aim for sentences that can be lifted without losing meaning: “Contains 20g whey isolate per serving and is Informed-Sport certified.” “Fits 60 x 60cm cabinets and runs at 44dB on eco mode.” “Suitable for eczema-prone skin. Fragrance-free and dermatologist tested.” Not marketing poetry. Evidence. **3) Treat “AI visibility” as a measurable channel** Lmo7 runs audits across multiple models (ChatGPT, Claude, Gemini, Grok) because visibility is not uniform. The same brand can look strong in one model and absent in another. If you are serious about agentic commerce, you need a baseline and a loop: - measure mentions and citations - ship changes - retest prompts - track which pages get pulled as sources Quick checklist for eCommerce teams this week A) Rewrite your top 10 product pages so the first 400 to 600 words are dense with buyer-critical facts. B) Add a “Best for / Not for” section written in plain language. C) Turn key claims into explicit, quotable sentences. D) Split “mega guides” into modular pages that each win a specific intent. E) Track which URLs assistants cite today, then work backwards into why. Get in touch if you want help with this!