Definition
Passage ranking is the scoring step that puts individual chunks of text in order, so the best few paragraphs get handed to a model rather than whole pages. Reranking is the closest neighbour and sits one step later: passage ranking produces the shortlist, reranking reorders that shortlist with a slower, more accurate model.
At a glance
- CategoriesTechnical
- Related fieldsContent, SEO
- DifficultyAdvanced
At a glance
- The unit being scored is a passage, roughly a section, not a URL.
- One page can contribute a single strong passage and have the rest ignored.
- Passages are scored against a sub-question, not the question a person typed.
Scoring paragraphs instead of pages
Your page gets split before it is judged. The system cuts it into chunks, embeds each one, and scores each against the question it is working on.
That question is usually not the one that was typed. A request gets broken into several background searches by query fan-out, and your passage is scored against one piece of it, so the match is narrower than you would expect.
Scoring reads the question and the passage together rather than counting shared words, the approach that became standard with the BERT paper. A passage answering in different vocabulary scores well, and a passage stuffed with the exact phrase can score badly. The top few go forward. Everything else was never in the conversation.
Your strongest section carries your weakest page
Page-level thinking hides the thing that matters. A page is not good or bad here. Its sections are, one at a time.
Take a long buyer’s guide. One section answers “how do I know what size to order” cleanly, under its own heading. The rest wanders through brand history and filler. That section can get cited repeatedly while the other twenty are never retrieved, and your analytics will not say which part did it.
The reverse costs more. A useful page written as one continuous argument, each paragraph leaning on the last, chunks into fragments that answer nothing. Good content, unretrievable shape. The lever is not page length but whether each part survives being lifted out, which is the discipline behind AI search ranking generally.
Cut your pages into liftable units
Take your three highest-value pages and run this.
- Give every section a heading that states its subject. Headings mark where chunks are likely to start and tell the scorer what the block is about.
- Open each section with the claim, not the wind-up. The first sentence should answer the implied question outright. Background comes after, if at all.
- Delete the back-references. “As we saw above” breaks a passage the moment it is pulled out alone.
- Keep one subject per section. A block covering sizing, shipping and returns scores mid on all three and wins none.
- Read each section out of order. If it needs the one above it, restate the missing context in a line.
Frequently asked questions
-
Is passage ranking the same as reranking?
Two stages of one pipeline. Passage ranking scores a large pool of candidates quickly to produce a shortlist. Reranking reorders that shortlist with a heavier model reading each candidate closely. Fast and wide, then slow and precise.
-
Does a long page hurt my chances of being cited?
Length itself is neutral and structure decides it. A long page of clean, self-contained sections gives a system many chances to match. One written as an unbroken argument gives it fragments that answer nothing. Plenty of long pages get cited constantly.
-
How do I find out which passage of mine got used?
You mostly cannot, and anyone claiming otherwise is guessing. Assistants cite a URL, not an offset into it. Read the answer’s wording against your page instead: the source section is usually obvious, and then you write more sections shaped like it.
-
Should I split my pages into lots of shorter ones?
Usually not. Splitting a strong page into ten thin ones scatters the topic, and thin pages struggle on every other signal. Fix the internal structure first: headings, self-contained sections, one subject each. Split only when a page covers two unrelated jobs.