Passage retrieval
Passage retrieval is when a search system picks out one section of a page that answers a question, instead of treating the whole page as one.
Passage retrieval is when a search system or AI platform finds the one section of a page that answers a question, instead of judging the page as a whole. A long guide might cover ten things. If one paragraph in the middle answers the question well, that paragraph can be what gets used.
Google describes its own version as a passage ranking system that identifies individual sections of a page to understand how relevant the page is. Systems that use RAG work in a similar way: they split pages into chunks, find the chunks that match the question, and hand those to the model.
Why it matters
It means each section of your page competes on its own. A good page with the answer buried in the seventh paragraph, under a vague heading, is harder to use than a page that answers the question near the top under a heading that says what it's about.
It also means a section has to make sense alone. If a paragraph says "it costs £12 a month" and "it" was named three paragraphs up, the passage loses the brand when it's lifted out.
An example
Fernhill Coffee's subscription page is long. It covers the story of the business, how the subscription works, delivery areas and prices. Someone asks an AI platform, "Does Fernhill deliver to Ireland?"
The platform doesn't need the story. It looks for the passage about delivery. If that section says, "Fernhill Coffee delivers across the UK and Ireland, with free delivery over £25," it can be used as it stands. If it says, "We ship to most places, see the table below," and the table is an image, the platform has nothing to quote.
How to write for it
- Give each section a heading that says what it answers
- Open each section with the answer, then explain
- Name things in full, so a passage makes sense lifted out of the page
- Write facts as text, and keep them out of images
Page checks in Recommendations look at some of this for a page and a prompt: whether the prompt appears in a heading, and whether the page opens with an answer.