Passage Ranking

Google lists passage ranking among its ranking systems and describes it as an AI system it uses to identify individual sections or “passages” of a web page to better understand how relevant the page is to a search^[1]. By that description it is a way of judging a page, so the thing that appears in the results is still the page.

Google first announced it in October 2020. It said very specific searches can be the hardest to get right, because the single sentence that answers a question may be buried deep in a web page, and that understanding passages in addition to the relevance of the overall page would find it. It said the technology would improve 7 percent of search queries across all languages as it rolled out globally^[2]. That figure was Google’s expectation at the rollout in 2020, not a measurement of what the system does now. Its example highlighted a featured snippet, but Google said the improvement applies to how it ranks web pages overall.

The research behind the idea is public. A 2019 paper by Nogueira and Cho described a re-implementation of BERT for query-based passage re-ranking that, at the time, was the state of the art on the TREC-CAR dataset and the top entry on the MS MARCO passage retrieval leaderboard^[3]. That paper is research, not a description of Google’s system.

It does not mean a site should chop its pages up. Google’s guide to generative AI features says there is no requirement to break content into tiny pieces for AI to better understand it, that its systems can understand the nuance of multiple topics on a page and show the relevant piece to users, and that there is no ideal page length^[4]. The unit a writer controls is a section that reads on its own inside a page worth reading.

Access still comes first, and it differs by engine. In QueryRev’s study, the crawlers it counts as AI training crawlers are blocked by name on 7.9% of the top sites read on average, and the crawlers it counts as AI search crawlers on 4.8%^[5]. A passage on a page that a crawler is told not to fetch cannot be read by that crawler, however well it is written.

Search Console reports on pages and queries, not passages. The passage is something you map yourself: for a long page, list the queries it earns impressions for, and note which section answers each one.

Look for the narrow queries. A page that earns impressions for a very specific question that only one deep section answers is a page where a passage is doing the work. Count how many of a page’s queries map to sections below the first screen.

Then judge each section on its own. Read it without the rest of the page: does it name its subject, does it give the answer in its first sentences, and does it make sense to a reader who arrives at it directly?

Illustrative: an invented site, to show the method.

A garden-supplies site has a long guide to choosing a hose, with six sections. A reader searches “what hose diameter for a low pressure tap”. Only the fifth section, “Diameter and water pressure”, answers it, in a short table and two plain sentences. The guide is not about low pressure taps: it is about hoses.

In Search Console the guide earns impressions for that query and for several like it. The site keeps the section complete, makes its heading say the question, and checks that the sentences under the heading name the subject rather than saying “as above”. The page is the result; the fifth section is why.