Your page gets quoted at the passage, not the page

Retrieval systems do not quote your page. They quote a passage from it, read alongside the text around that passage, and the difference decides whether you get cited.

Google published a guide in May 2026 called Optimising for generative AI features, and it is the most useful thing written on this subject partly because of what it rules out. It says llms.txt files are not used by Google Search. It says structured data is not required for generative AI search. It says, in plain words, that there is no requirement to break your content into tiny pieces. A cottage industry has grown up selling all three. Google's own documentation says none of it is the mechanism.

The actual mechanism is retrieval-augmented generation with grounding, plus query fan-out, which Google describes as running concurrent related queries off the one thing a person typed. Your page is not competing for a query. It is competing for a handful of sub-questions the system decided the query implied.

What self-contained really means

A passage is retrievable when it answers one question completely, in its own words, without needing the sentence before it.

Most business writing fails this by habit. The paragraph opens with This means that, or As we mentioned above, or a pronoun with no visible antecedent. Read on the page, it flows. Pulled out on its own, it is a fragment with a dangling reference, and it does not survive extraction.

Compare two versions of the same idea.

Weak: This is why we always recommend it for the coastal jobs. It handles the exposure far better than the alternatives, and the cost difference is minor over the life of the asset.

Retrievable: For coastal sites within about a kilometre of surf, hot dip galvanising to AS/NZS 4680 outlasts a standard paint system by a wide margin, and the cost difference is usually recovered inside the first maintenance cycle.

Same claim. The second one names the standard, names the condition, and stands alone. It also happens to be more useful to a human skimming, which is the part people forget.

Why you still do not chunk the page

Here is the bit that confuses people. Retrieval happens at the passage level, so surely you should publish passages?

No, because the surrounding context is part of what gets read, and because a page of disconnected fragments is a bad page for the human who lands on it after the citation. Google is explicit that fragmenting content is not required. What you want is a coherent page made of paragraphs that each hold together on their own. That is not a new writing skill. It is what good technical writing has always looked like.

Practically: one idea per paragraph, the claim in the first sentence, the qualification in the second, the specifics in the third. Subheadings that are plain questions or plain noun phrases so the section boundary is unambiguous. No paragraph that opens by referring backwards.

The eligibility gate nobody checks

Google's guidance says the page must be indexed and eligible to appear with a snippet.

Read that twice if your site is running any snippet controls. If you have nosnippet on a template, or max-snippet set low, or data-nosnippet wrapped around a section someone once decided was too valuable to give away, you have removed yourself from generative eligibility on those pages. Plenty of sites did this years ago to protect content and forgot.

Check it tonight. View source on your ten most important pages, search for nosnippet and max-snippet, and check the robots meta tag and the X-Robots-Tag header. This is a ten-minute job with a real chance of finding something expensive.

Being cited is now worth roughly double

Seer Interactive tracked 53 brands, 5.47 million queries and 2.43 billion impressions from January 2025 to February 2026. On results pages with an AI Overview present, organic click-through ran at about 2.1% when the brand's page was cited inside the Overview, and about 0.9% when it was not. On results pages with no AI Overview at all, about 3.3%.

Two things follow. Being cited more than doubles your click rate on the same query. And an AI Overview appearing at all still costs you against the clean results page, so citation is damage control rather than a windfall. Ahrefs measured a 58% click reduction for the top-ranking result when an Overview appears, which is the same story from the other side.

The recovery is real too. That click rate with an Overview present bottomed at 1.3% in December 2025 and had climbed back to 2.4% by February 2026. Whatever Google did to the format between those dates gave some clicks back.

Writing for fan-out without writing filler

Query fan-out means one page can be retrieved for several sub-questions at once, which is an argument for covering the adjacent questions on the same page. It is not an argument for padding.

The test is whether each section answers a question a real buyer asks out loud. For a commercial roofing page that is: what does it cost per square metre, how long does it take, what happens to the tenants during works, what is the warranty, who carries the defects risk, what goes wrong on older concrete decks. Six real sub-questions, six retrievable passages, one coherent page.

Compare that to the version that opens with three paragraphs on the history of roofing. That section will never be retrieved for anything, and it pushes the passages that could be down the page.

The honest limits

Passage-level writing does not fix a page nobody should read. If the content is commodity, cleanly structured commodity still loses.

You also cannot measure this well yet. There is no report that tells you which passage got quoted. You can see whether you appear in Overviews for tracked queries, and you can watch the click-rate split, but the attribution stops there. Anyone selling you passage-level citation analytics is estimating.

And the volume side is sobering. AI Mode was 0.34% of searches in the January to April 2026 window. The pressure is coming from AI Overviews on ordinary results pages, not from the chat surface everyone is planning around.

Write the paragraph so it survives being lifted out. That is the whole discipline.