Writing for Retrieval

Track 2 · Lesson 3 of 6 · about 15 minutes

Writing for retrieval

Lesson one established that selection happens at passage level. This lesson is about what a winning passage looks like — and the uncomfortable news is that it is mostly old-fashioned good writing, enforced more strictly than human readers ever enforced it.

Self-containment is the master property

A retrieval system lifts a chunk out of your page and hands it to a model with limited surrounding context. A paragraph that only makes sense after the three above it is nearly useless in that pipeline, however good the full page is. The test: could this paragraph be read alone, by someone who has not seen the rest, and still deliver its claim? If not, it will not be the passage that gets quoted.

Practically, that means fewer pronouns doing long-distance work (“this approach”, “these results” — which approach? which results?), key terms restated rather than referred back to, and one idea per paragraph rather than three ideas braided together.

Answer first, then expand

The classic journalism structure — conclusion at the top, detail below — turns out to be retrieval-optimal. A section that opens with its answer in forty words gives the retriever a dense, extractable target. A section that builds suspense for five paragraphs gives it nothing until the end, and retrieval rarely waits.

Headings do double duty here: an H2 phrased as the question a searcher asks, followed immediately by its answer, is the shape both featured snippets and AI citations tend to lift.

Specificity earns quotation

Compare “site speed matters for SEO” with “INP replaced First Input Delay as a Core Web Vital in March 2024.” The first is filler every page on the internet contains; there is no reason to cite any particular source for it. The second is checkable, dated and precise — a sentence a careful system can safely repeat with attribution. Numbers with sources, dates, named mechanisms, honest scope (“in our tests”, “as of mid-2026”) all raise a passage’s quotability.

The inverse also holds: a confident claim that is checkably wrong is worse than no claim, because systems that verify sources learn to discount you.

What this does not mean

It does not mean robotic prose, question-stuffed headings on every section, or deleting narrative from your writing. Pages are still read by people, and a page that reads like a FAQ dump loses the human reader the citation was supposed to bring you. The craft is writing that works at both levels: sections that flow for a reader, each anchored by a passage that stands alone for a machine.

Going deeper: the GEO and AEO guide · practise on Would AI cite this?