Query Decomposition

Query decomposition is the AI-era technique of splitting a complex question into several sub-queries, answering each, then synthesizing a final answer. It is what Google AI Overviews and Perplexity do under the hood, and it changes what you should write.

The user asks one thing. The model asks five. The page that answers all five is the one that gets cited.

How query decomposition works

• The original query is rewritten into 3–10 sub-queries that, together, would answer the original
• Each sub-query is run through the retrieval system separately
• The best passages for each sub-query are collected
• The model writes a single answer that synthesizes them, with citations

Example

• Original: “Is Kyoto a good destination for a family with young kids in October?”
• Sub-queries: “weather in Kyoto October”, “family activities Kyoto with kids”, “Kyoto with toddlers stroller-friendly”, “Kyoto October crowds lines”
• The model retrieves passages answering each, then writes a single answer citing all of them

How to optimize for query decomposition

• Cover the sub-questions in one place. A page that answers all five sub-queries is more likely to be cited than five separate pages
• Use H2s that match the sub-queries. Models chunk pages by heading; clear H2s help the model find the right passage for each sub-query
• Use schema markup FAQ. Explicitly map your content to the questions
• Anticipate the decomposition. For any topic, ask yourself: what are the 3–10 sub-questions an expert would break this into?
• Build topic clusters. A pillar can serve as the synthesizer; the cluster pages serve the sub-queries