The passage-first content model: what's real, what's ritual

What is the passage-first content model?

Passage-first means writing so an AI can lift a clean, self-contained answer straight off your page. AI search tools cite passages, the small sections of a page, not whole pages. So the thing you optimise is the passage, not the article. Write each section to stand on its own and you hand the machine something clean to quote.

A passage is just a few sentences that fully answer one question, with no loose ends. "What does our onboarding include?" answered in three plain sentences is a passage. The same answer split across two paragraphs that each need the other to make sense is not.

Three confirmed mechanisms sit underneath the idea, and we'll grade each in turn: Google ranks individual passages, AI search breaks one question into many, and the answer you see is lifted from real sentences rather than written from memory. That foundation is documented and supported. The writing conventions built on top of it are where the folklore creeps in.

Is it true that AI reads passages, not pages?

Yes. This is the solid ground, and it's the reason the whole model exists.

Google launched passage ranking on 10 February 2021, for US English searches, affecting about 7% of queries. One nuance gets lost constantly: Google still indexes whole pages. A passage is an extra ranking signal, not a separate index. Danny Sullivan from Google had to correct the early name, "passage indexing," because it made people think otherwise. So a single strong section can carry a page that's weak elsewhere.

Next, query fan-out. At Google I/O 2025, Liz Reid described how AI Mode takes your question, breaks it into subtopics, and fires off many searches at once. You ask one thing; the system asks ten. Each of those mini-searches looks for a passage that answers it.

Then grounding. When an AI shows you an answer, it isn't writing freehand from memory. It pulls real sentences from real pages and stitches them together. Dan Petrovic's firm DEJAN reverse-engineered how Google's Gemini does this and found that sentences appearing early and standing alone are far more likely to be pulled, and that "density beats length." Treat this as illustrative rather than settled: it's one firm's work, not peer-reviewed, and the dataset isn't public.

Put those together and the takeaway is simple. The machine quotes sentences, not paragraphs. Writing that survives being lifted out of context is writing that gets cited.

Where does the "40 to 60 word answer block" rule actually come from?

From one study. In 2018, Ghergich and Co. analysed roughly 1.4 million featured snippets with SEMrush and found the boxed answers ran about 40 to 50 words. That's a measurement of what fit inside Google's snippet box, not a rule about what AI retrieves. Google has never stated a minimum or ideal length.

This matters because almost every AEO guide repeats the number without telling you where it came from. Backlinko says 40 to 60. Moz found 45 to 97. Portent found 40 to 55, and Portent's own author has said he doubts Google is looking for any particular word count. The figure describes how answers were displayed in a search box between 2018 and 2020, before AI search existed. It was never a retrieval law.

A short answer is easier to quote, so the instinct isn't wrong. But 61 words won't get you dropped and 59 won't get you picked. We removed the word-count rule from how we describe our own AEO work, because it doesn't survive scrutiny from a senior SEO. Aim for tight. Don't count to 60 and stop mid-thought.

Does the writer even control the chunks?

No, not fully. You influence where the boundaries fall. The retrieval system decides them.

There's a real disagreement at the centre of this topic. Danny Sullivan from Google says don't chop your content into bite-size pieces for machines, and that there's nothing special creators need to do. Meanwhile Mike King at iPullRank and Lily Ray at Amsive have shown that structure measurably changes what gets retrieved. King's testing found that splitting a paragraph covering two topics improved how well each part matched a query, and that adding a heading helped another. As Lily Ray puts it, "they, not you, decide how to slice your content."

Both sides are right, which is the honest synthesis. Structure helps humans and machines at the same time. But "writing for chunks" as a mechanical ritual is partly superstition, because you don't own the cut. King's sharper point is that vague "just write naturally" advice is how Google keeps marketers as its unpaid workforce, doing the structuring work without a manual.

Write clearly, structure honestly, and stop pretending you control the slicing.

You're setting up the content so the machine's cut lands in a sensible place. That's influence, not control.

So what actually helps?

Self-contained passages, answer-first writing, and plain sentences. These have the strongest evidence behind them. Schema does not.

Three conventions earn their place:

• Self-contained passages, the strongest-supported one. Each section should make sense on its own. Don't open a sentence with "It" or "This" when it only works if you've read the paragraph before. Name the thing. Petrovic, Lily Ray, and Mike King all land in the same place here.
• Answer-first, also called the inverted pyramid. Lead with the answer, then elaborate. Journalists have written this way for a century. Petrovic's data shows early, standalone sentences get pulled most often.
• Plain, literal sentences. The clean literal sentence is exactly what an AI can quote. Clever, allusive prose reads beautifully and extracts terribly.

Here's the part we keep coming back to with clients. Plain English is better AEO, not a tax on your brand voice. The sentence a person understands fastest is the sentence the machine quotes. They aren't competing goals.

And the worry we hear most: won't this fragment my writing into bullet soup? No, if you get the rule right. Be tight per passage, long per page. You don't write less. You stack more real answers, each one clean. Concision at the passage level, coverage at the page level.

When is passage-first the wrong priority?

When the citations won't turn into pipeline, when your buyers aren't triggering AI answers, or when your best content is the kind that resists summarising.

Citations aren't clicks. Pew Research found that when Google shows an AI summary, people click a link inside that summary just 1% of the time (68,879 searches, July 2025). For bottom-of-funnel B2B pages, the ones meant to drive a demo or a call, weigh the visibility against the pipeline before you optimise hard for a quote nobody clicks through.

Then there's your query mix. If your buyers' searches rarely trigger AI Overviews, passage-first sits lower on the list than other work you could be doing. Check before you commit a quarter to it.

Last, some content should resist summarising, and that's the point. Proprietary data and a genuine opinion are hard to compress, which is exactly why they get cited and remembered. iPullRank and others argue AI rewards content it can't fully flatten. Don't atomise a strong argument into fragments chasing a word count. The argument is the asset.

The honest version

The mechanism is worth building for. AI search reads and cites passages, that's confirmed and it isn't going away. The rituals around it, the magic word count, the schema-for-snippets myth, the semantic-triple incantations, mostly aren't.

So grade before you obey. Ask of any AEO tactic: is this documented, supported, or folklore? Keep the documented and the supported. Drop the folklore. Then write plain, self-contained answers, because that's good writing first and good retrieval second, and the two were never at odds.