Passage Ranking: Definition and Stakes for AI Visibility

Most people still picture search as a contest between pages. It has not been that for a while. Since Google added a passage ranking system to its published list of ranking systems, the engine reads inside a page, isolates the sections that could answer the query, and scores those. Then generative engines industrialised the same idea: they cut documents into chunks and build their answer from the two or three they retrieve. The unit of visibility moved from the page to the passage. This page defines the term, separates it from the myth it gets confused with, explains the machinery underneath, and shows what a passage that actually gets picked looks like.

Passage ranking, defined in one line

Passage ranking is the search system that scores individual sections of a web page, rather than only the page as a whole, to judge how well that page answers a query. Google lists it among its ranking systems as an AI system used to identify individual sections, or "passages", of a web page in order to better understand how relevant the page is to a search. One strong paragraph deep inside a long page can now win a query the page overall would never have won.

That definition is not an interpretation. It comes straight from Google's own guide to its ranking systems (Google Search Central), where the passage ranking system sits alongside RankBrain and the helpful content signals. Two words in it matter. It is a system, permanently on, not an update that shipped once. And its job is to look inside the document, which is exactly what search spent two decades not doing.

The consequence lands on a page you already own. A long guide covering eight subtopics used to be judged as an average of those eight. Now section five can be pulled forward on its own merits while the other seven stay dormant. Your page does not rank. Your paragraphs do.

Passage ranking is not passage indexing

Google indexes pages, not passages. Your page enters the index as one document, and passage ranking then looks inside it so a specific section can carry the relevance signal for a specific query. The confusion comes from the original 2020 wording, which is precisely why the system ended up being named passage ranking.

The story is worth two lines because the misunderstanding still shapes bad advice. At its Search On event in October 2020, Google announced that it could understand individual passages inside a page rather than only the page as a whole, and the rollout for English queries in the United States followed in early 2021. Plenty of people read that as "Google now indexes passages separately", concluded that pages no longer mattered, and started chopping content into thin standalone URLs.

That was the wrong lesson, and it cost real traffic. Nothing about the index changed. Your page is still crawled, indexed and ranked as a page; what changed is that a section inside it can be the reason it ranks. Splitting a solid guide into fifteen shallow pages does not win you passages. It gives you fifteen weak pages.

How passage ranking actually works

It relies on language models that read a page in segments and score each segment against the query. That is the same architecture the research literature has used for years: passage re-ranking with BERT, dense passage retrieval, and benchmarks built at passage level. The engine compares your paragraph to the question, not your whole page to the topic.

The academic line is public. In 2019, Nogueira and Cho published Passage Re-ranking with BERT, showing a transformer could re-score candidate passages against a query far better than the keyword matching that came before. A year later, Karpukhin and colleagues published Dense Passage Retrieval (EMNLP, 2020), which retrieves passages by meaning rather than by shared words. And the benchmark the whole field competes on, Microsoft's MS MARCO, is a passage ranking benchmark. Not a document one.

For most of the past decade, the unit of retrieval in the research that powers modern search has been the passage. Google's passage ranking system is not a feature bolted onto search. It is search catching up with how its own machinery already thinks.

Why passages decide your AI visibility

AI answer engines work the same way, only more so. Retrieval-augmented systems cut documents into chunks, embed them, retrieve the few most relevant chunks, and compose the answer from those. Whether the surface is Google, ChatGPT or Perplexity, what gets selected is a passage, not a page. Passage-level writing is the one habit that pays in classic search and in AI answers at once.

This is where the term stops being Google trivia and starts deciding whether an AI names you. When an assistant answers with sources, it did not read your site: it retrieved a handful of chunks, a few hundred words each, and wrote from those. Anthropic's engineering write-up on contextual retrieval (Anthropic, 2024) is candid about the failure mode: a chunk cut away from its document loses the context that made it meaningful, and retrieval then misses it. Their fix is to re-attach context to each chunk before embedding. Yours is simpler and free: write passages that carry their own context in the first place.

The second reason to care comes from our own numbers rather than a vendor blog. Passages win precise, narrow questions, and that is where the search volume actually lives. Across the 4836 French keywords Cicero Studio has analyzed, 34% get fewer than 100 monthly searches, and on the 4253 French keywords where we could measure a volume, the median is 260 searches/month. The long tail is not a footnote to the strategy. It is the strategy. Those small, specific questions are exactly what a single well-built passage can win, one at a time, in aggregate, and exactly what head-term thinking writes off as too small to bother with. Our French deep-dive on la longue traîne SEO unpacks the arithmetic.

What a passage-ready section looks like

A passage-ready section survives being cut out of the page. It names its own subject instead of pointing back with a pronoun, answers one question completely, repeats just enough context to stand alone, and carries its source inside itself. If a paragraph only makes sense in sequence, it will not be picked.

The test we run on our own drafts is blunt: copy one paragraph into a blank document, hand it to someone who has not read the page, and see whether it still means anything. Most paragraphs fail. They open with "this approach", "as we saw above", "the second one", ropes tying them to a context the engine is about to cut away.

Here is the same paragraph before and after that treatment, taken from a rewrite we ran on one of our own pages:

Same idea, same length. The first version dies the moment it is cut out of the page: "this", "the second category", "as we saw earlier" all point at something the engine no longer holds. The second names its subject, states the mechanism and closes the loop on its own. That is the whole difference between a paragraph that gets quoted and one that never leaves your site.

Headings do disproportionate work here: they are the most reliable boundary an engine has for deciding where one passage ends and the next begins. Vague headings mean badly cut chunks, and badly cut chunks are, in our experience, the most common reason a genuinely good article never gets quoted anywhere. Two of our French deep-dives cover the mechanics: one on how AI chunking reads your content, the other on AI agents that search passage by passage.

Passage ranking, snippets, AEO and GEO

Passage ranking is the mechanism: the engine scores sections instead of whole pages. Featured snippets and AI Overviews are surfaces where the winning passage gets shown. AEO and GEO are the disciplines of optimizing for those surfaces. One is plumbing, the others are strategy.

These four terms get used as if they were rivals. They are not: they sit at different levels of the same stack.

We define each on its own page: AEO, answer engine optimization, GEO and LLMO, with the side-by-side in our SEO, AEO and GEO comparison. Passage ranking sits underneath all of them, which is why it is worth understanding first.

The limits of passage ranking

Let me be straight about what this concept will not do for you, partly because that honesty is itself the kind of signal engines reward, and partly because half the advice written about passage ranking oversells it.

Underneath the vocabulary, passage ranking asks for something old-fashioned: write in complete, honest, self-standing units, and put the source next to the claim. It happens to be what a machine needs. It is also, not coincidentally, what a reader in a hurry needs.

Test your own page in ten minutes

Take the question your best page is supposed to answer. Search it, and look at what got lifted: a passage from your page, a passage from someone else's, or nothing of yours at all. Then find the paragraph on your page that should have won and read it out of context. Nine times out of ten, you will hear immediately why it lost.

Do it on a page you are proud of, because that is where it stings and where it teaches. We ran it on one of our own pillar pages last winter and watched it lose to a competitor's plain FAQ answer: our best paragraph opened with "this method" and never said what the method was. The engine had nothing self-contained to lift. Across the 1207 SEO/GEO audits produced by Cicero Studio, we found that same pattern repeating: the content is there, the expertise is there, and it is packaged so that no machine can cut it cleanly. Good news, in fact, because the fix lives on pages you already own rather than in a fresh production plan.

Where Cicero Studio fits

Cicero Studio treats the passage as the working unit inside one method: a GEO audit that maps your visibility on Google and in AI answers, an AI-augmented editorial production reviewed by humans, and automated semantic internal linking. It starts with a free audit, so the diagnosis comes before anything gets written.

At Cicero Studio, that means every section we ship is built to be lifted out. The audit tells us which questions are winnable, production writes standalone passages against those questions rather than essays around a theme, and the link mesh keeps each passage anchored to its cluster. That is the method hook: GEO audit, editorial production, automated semantic meshing, run as one loop instead of three disconnected services.

We run on ourselves the test we just handed you. The 502 articles published on cicero.studio (264 FR, 238 EN) are our proving ground: each is written passage by passage, then measured on whether those passages get picked. That is what we mean by agency-quality work, software-grade productivity. The French treatment of the model lives on our agence GEO pillar; our English breakdown of a GEO audit shows what we check.

Alexis Dollé

CEO & Founder of Cicero Studio

A growth specialist and content strategy consultant, I founded Cicero to help businesses build durable organic visibility, on Google as in AI answers. Day to day, I run our clients' audits and editorial production: we put AI to work for production, never in place of expertise. Every piece is built to convert, not just to exist.

LinkedIn →

Resources to go further

We document the method in the open, because published work with its sources beats any sales deck. Each link below digs into one piece of the passage puzzle. Several are in French, our home market, and are flagged as such: