Late chunking
Late chunking is a way of splitting a document for AI search that reads the whole document first, so each piece keeps its surrounding context.
Late chunking is a way of preparing a long document for AI search so each piece of it keeps some of the context around it. The system reads the whole document first, then splits it into pieces, called chunks. The usual way does it the other way round: split first, then read each piece on its own.
It's a technical detail inside search systems. It's worth knowing because it shows why a passage that only makes sense with the page around it can lose out.
Why documents get chunked
AI search systems rarely compare a question with whole pages. They cut pages into chunks of a few sentences or paragraphs, turn each chunk into a list of numbers that captures its meaning, called an embedding, and then look for the chunks closest to the question.
The trouble is that a chunk read on its own forgets where it came from. A paragraph that says "It ships in two days and costs £12 a month" doesn't say what "it" is.
What late chunking changes
With late chunking, the model reads the whole document in one go, so every word is understood in light of the rest. Only then is it split up. The chunk about "£12 a month" keeps some sense that "it" means Fernhill Coffee's subscription.
The method was described by researchers at Jina AI in a 2024 paper. Which systems use it isn't public, and the AI platforms don't say how they chunk pages.
What it means for your pages
You don't control how a platform chunks your page, so don't rely on it to fix things. Write so each section stands on its own: name the product, don't lean on "it" or "this" from three paragraphs up, and start each section with the point.
For example, a section headed "Price" that says "Fernhill's coffee subscription costs £12 a month" will be understood whichever way the page is cut.