Prose enters an AI answer by passing three gates. It gets retrieved when its meaning matches the query. It survives reranking, the scoring pass that keeps only the most useful retrieved passages. And it gets credited when the model can lift a clear, attributable claim from it. Writing style decides all three, and the style that wins is the one this guide is written in: answer first, subject named, claims specific.
The extractable passage: anatomy and a worked example
A passage built for extraction has four properties. The answer sits in the first sentence, because the chunk may get cut before your conclusion arrives. The subject is named inside the passage, with no pronoun pointing at a previous paragraph the model never sees. Each sentence carries one claim. And the claims are specific enough to be worth quoting.
Before: "When it comes to choosing a backup solution, there are many factors to consider, and the right choice really depends on your specific needs. Our platform offers industry-leading flexibility." Nothing there survives extraction; there is no subject, no claim, no number.
After: "Acme Backup restores a 1 TB server image in under 40 minutes on a gigabit connection, and its lowest tier includes hourly snapshots, which most competitors reserve for enterprise plans." One brand, two checkable claims, a number, and a comparison, all inside a single liftable passage.
Evidence density is measurably rewarded
The GEO paper (Aggarwal et al., KDD 2024), the first systematic benchmark of content changes against generative engines, reports that content optimizations can lift visibility in generative answers by up to 40%, with effectiveness varying by domain; the tactics tested include adding statistics, quotations, and cited sources.GEO: Generative Engine Optimization, Aggarwal et al. One benchmark on one generation of engines, so treat the 40% as an existence proof that writing changes visibility, and expect the exact figure to shift as engines change. The mechanical reason is in the gates above: a reranker scores passages by usefulness for answering, and a passage carrying a number, a named source, or a direct quote gives the model something it can assert and attribute.
This is also the difference between being the source of a fact and paraphrasing someone else's. If your page says "studies show most users abandon slow sites", the citable fact belongs to whoever ran the study, and an engine has every reason to cite them instead of you. If your page says "in our logs across 1,400 client sites, [number: verify] of visits abandoned pages slower than three seconds", you are the origin, and citing the fact means citing you. Evidence, Expertise, and Original Research covers becoming that origin deliberately.
The formats that keep winning
Certain shapes of content map directly onto the prompt intents from Prompt and Topic Research, which is why they keep showing up in answers: comparisons with real criteria, definitions stated plainly in the first sentence, numbered steps for how-to intent, spec and pricing tables in text, and FAQs whose answers actually answer. The format is doing the same job in every case, matching the shape of the question and handing the model a ready passage.
What to stop doing
Narrative windups: the three paragraphs of scene-setting before the point. Human readers skim past them; a chunker may keep them and cut your conclusion instead. Empty hedges: "results may vary" and "it depends on your needs" state nothing an engine can use. Say what it depends on, or make the call. Hedging a claim honestly, with the reason attached, is fine writing; hedging to avoid having a position produces prose with no extractable claim in it.
The subtler failure is formatting cargo cult: content that apes the shape of retrievable writing, the bullet lists, the FAQ blocks, the bolded question headings, while saying nothing an engine would quote. Structure without claims is still nothing. Google's guidance points the same direction from the platform side: its systems handle synonyms and phrasing variation, so rewriting content "for AI" in some special register is unnecessary, and the standard that matters is helpful, reliable, people-first content.Google's AI features guide (mythbusting section) and its people-first content guidance, which defines the E-E-A-T standard: experience, expertise, authoritativeness, trust. There is no conflict between writing for extraction and writing for people. Clear, specific, well-attributed prose is both; the cargo cult is neither.
Editing existing content without gutting it
Most sites need edits, and the edit pass is mechanical enough to delegate. Per section: move the answer into the first sentence, replace pronouns whose referents live outside the section with the actual subject, attach a number or named source to the load-bearing claim, and cut the windup. Keep the voice, the narrative, and the human texture between those points. The goal is a page a person still enjoys reading where every section, lifted alone, says who it is about and what it claims. If an edit makes the page worse for a human reader, you have crossed from optimization into cargo cult; back up.