Grounding Budget

The grounding budget is the fixed amount of retrieved content an AI system allocates per query when constructing a response. Research from Dan Petrovic at DEJAN analyzing over 7,000 queries found that Google’s Gemini allocates approximately 2,000 words of grounding content per response, distributed across multiple sources. This hard ceiling means that most of the content on the web that could answer a query never makes it into the AI response.

How the Grounding Budget Works

When a user asks a question, the AI retrieval system searches for relevant passages, ranks them by relevance and authority, and selects content to fill the grounding budget. The budget is split across sources: a typical response might allocate 400 to 600 words from the top source, 200 to 400 from the second, and diminishing amounts from additional sources. Sources ranked lower in the retrieval pipeline receive smaller allocations or are excluded entirely.

Dan Petrovic’s research found that approximately 32% of any given page’s content survives the grounding filter on average. Pages under 1,000 words achieve roughly 61% coverage while pages over 3,000 words achieve just 13% coverage. This inverse relationship between page length and survival rate has a direct implication: shorter, denser content with front-loaded answers outperforms long-form content that buries key propositions deep in the text.

Implications for Content Strategy

Front-load your answer. The inverted pyramid structure is not optional. AI systems exhibit lead bias and disproportionately extract from opening paragraphs. If your key proposition appears in paragraph five, it will likely fall outside the grounding budget.
Increase atomic density. Every word that is not a verifiable, extractable claim dilutes your content’s chance of surviving the budget. Marketing language, transitional phrases, and restated conclusions consume word count without contributing citable atoms.
Optimize per section, not per page. Each H2 or H3 section is evaluated independently during retrieval. A 3,000-word page with one strong section and four weak sections loses 80% of its grounding budget to passages that will not earn citations.

Platform Variations

The grounding budget varies by platform and query type. Google AI Overviews tend to use shorter grounding windows because they appear inline with search results. ChatGPT and Claude allocate larger budgets for complex, multi-part queries. Perplexity shows its sources explicitly and tends to pull from more sources with smaller allocations per source. These differences mean the same content may survive grounding on one platform but get cut on another.

For the complete content optimization framework, see the Generative Engine Optimization guide.