Chunk retrieval
Chunk retrieval explained
This is why page level thinking fails in AI search. Your page does not get retrieved. A section of it does. If your best answer sits in paragraph nine after a preamble, and depends on context established in paragraph two, it will be pulled out with the context missing and will lose to a cleaner competing chunk.
Writing for chunks means self-contained sections: a heading that states the question, an answer that stands alone, entities named in full rather than as it or they, and numbers with units and dates attached inside the same block.
It also means shorter pages sometimes beat long ones, because a focused page produces cleaner chunks.
Chunk retrieval is the step where a system splits pages into passages, indexes those passages, and pulls the individual chunks that best match a query rather than whole documents.
One definition is worth more than ten opinions
Every term here links to the service that owns the work and the tool that measures it.
Three things you can test this week
None of these need a tool or a vendor. They take an afternoon and they tell you whether chunk retrieval is actually your problem.
Our own measurement method, with its definitions and stated limits, is published on the method page.
The definition is the easy part
Knowing what a term means takes five minutes. Knowing whether it is your problem takes an afternoon, and knowing what to do about it takes a method.
Every glossary page here ends with three checks you can run yourself and the primary sources behind them, because a definition you cannot act on is trivia.
Primary sources
Primary documentation for the mechanics described on this page. All published by the organisations that operate the systems in question.
Twenty buying prompts, four engines, you against three named competitors. Report in one working day. No call needed.
Chunk retrieval: common questions
How long should a chunk be?
Roughly one hundred to three hundred words that make sense read alone.
Do long guides still work?
Yes if every section is self-contained. They fail when sections depend on earlier context.
What is the most common chunking mistake?
Pronouns. A chunk that says it costs less than the alternative names nothing and gets skipped.
See what the engines say about you
Twenty buying prompts, four engines, your brand against three named competitors. Report in one working day.
- 20 buying prompts, your category
- ChatGPT, Gemini, Perplexity, AI Overviews
- You against three named competitors
- Report in one working day, yours to keep