The test every sentence has to pass, the structure that makes passing easy, and rewrites you can copy.
Content AI will quote is content made of self-contained statements: each key sentence names its subject, makes one specific claim, and still reads as true and complete when lifted out of the page with no surrounding context.
An assistant does not read your page the way a visitor does. It retrieves it, scans it for a passage that answers the question in front of it, lifts that passage, and moves on. Everything about writing for AI search follows from one fact: the unit of value is no longer the page. It is the sentence.
Three properties, and a sentence needs all of them.
Most marketing copy fails the first property almost entirely. It is written to be read in order, top to bottom, with each sentence leaning on the one before. That is good writing for a human and nearly useless for an extractor, because the extractor never reads in order.
The extraction test takes about ten minutes per page and needs nothing but the page itself.
If you cannot find five candidate sentences at all, that is the finding. The page has content but no answers in it.
| Before | Why it fails | After |
|---|---|---|
| “We’re passionate about helping teams do their best work.” | No subject a model can name, no claim, nothing checkable | “Acme is project management software for engineering teams of 20 to 200 people.” |
| “Pricing is flexible and tailored to your needs.” | Hedged; tells the buyer nothing | “Acme costs £12 per user per month, billed annually, with a 14-day free trial.” |
| “It integrates with the tools you already use.” | “It” is unresolved once lifted; no tools named | “Acme integrates natively with GitHub, Jira and Slack.” |
| “Studies show this approach dramatically improves outcomes.” | Unattributed; citing it would mean asserting something unsupported | Either name the study, its publisher and date inline — or delete the sentence. |
| “Unlike some competitors, we take a different approach.” | Comparison with no named side | “Acme stores data in the UK; Beta and Gamma store it in the US.” |
Acme, Beta and Gamma are placeholders. The pattern is the point: every “after” sentence would still be accurate and useful if it were the only sentence a buyer ever saw about you — which, inside an assistant’s answer, it may be.
The heading is the retrieval target. When an assistant is looking for an answer to “how long does onboarding take”, a heading that says exactly that is the closest possible match. A heading that says “Getting started with us” asks the model to infer.
The most common failure is burying the answer. The page does answer the question — four paragraphs after the heading, behind background nobody asked for. An extractor that reads the first passage under a heading and finds preamble will often move on to a competitor whose first passage is the answer. Move the answer up. Everything else can stay.
Yes, wherever the content is genuinely structured. A comparison written as prose forces the extractor to reconstruct the structure you already had in your head. A comparison written as a table hands it over intact.
Use a table when you are comparing two or more things across the same attributes: plans, products, approaches, before and after. Use a numbered list when order matters: steps, sequences, a checklist. Use bullets when items are parallel but unordered. Do not force structure onto content that is really an argument — an argument is prose, and should stay prose.
Every figure carries its source and its date, inline, in the same sentence. Not in a footnote, not in a references section at the bottom — the extractor lifts the sentence, not the footnote.
“Our customers see results fast” is unquotable. “Median onboarding time in 2026 was nine days” is quotable only if it is your measured figure and you can stand behind it. If you do not have the number, do not write a sentence that gestures at one. A model that repeats an unsupported claim on your behalf is doing you no favours, and the better assistants increasingly decline to.
This cuts against a lot of marketing instinct. Hedged, number-shaped language feels safe. For extraction it is the least safe thing on the page, because it is the thing most likely to be skipped.
Not the blog. Start with the pages that answer the questions a buyer asks when comparing vendors:
Five pages, rewritten properly, will usually do more than fifty new blog posts — because they are the pages that answer the questions that lead to a purchase.
One section, rebuilt from scratch, for a fictional product. The heading is the question a buyer types. The first sentence answers it completely. Everything after it is optional detail that a human might want and an extractor can ignore.
<h2>How long does Acme take to set up?</h2>
<p>Acme takes about a day to set up for a team of 50: one hour
to connect GitHub and Jira, the rest importing existing projects.</p>
<p>Larger teams usually stage the rollout by department. Migration
from Beta or Gamma is handled by a built-in importer, which keeps
issue history and assignees. SSO is configured separately by an
admin and takes under an hour with Okta or Azure AD.</p>Lift the first paragraph alone and it still works: named subject, specific claim, a number with context. The second paragraph serves the reader who wants more. Nothing in either depends on a sentence elsewhere on the page. (The figures belong to the fictional example; use your own measured ones.)
Quotable sentences scattered across a site can contradict each other, and a model that finds two different prices or two different descriptions of what you do will trust neither. Keep a short internal reference — one document — that holds the canonical version of the claims you most want repeated:
Every page that states one of these copies it from the reference, and every change starts there. It sounds bureaucratic. It is the cheapest insurance against the most common entity problem there is: a company that describes itself three different ways.
No. It usually makes it better, and this is the part worth holding onto if the idea of “writing for machines” is unappealing.
Everything that makes a sentence extractable — a named subject, one specific claim, a source for the number, the answer before the preamble — is also what makes it useful to a busy buyer skimming on a phone. The only thing that gets cut is the throat-clearing, and nobody was reading that anyway.
What you should not do is write stilted, keyword-stuffed text in the belief that assistants want it. They do not. They want the clearest available statement of a true thing. That is a writing standard, not a trick, and it is the same one good editors have always applied.
Pick the questions your rewritten pages answer and ask an assistant each one, several times, logged out. Record whether your page is cited, and which sentence was used. Then compare against the same questions asked before the rewrite.
Be patient and be honest about variance. Assistant output changes between runs, so a single appearance proves little and a single absence proves nothing. How to track citations without fooling yourself → And before any of this, make sure the page can actually be fetched — a perfectly quotable sentence behind a blocked crawler is still invisible. Check access first → If you want a second pair of eyes, the free check runs five buyer questions against your site and tells you which pages were cited instead.
Content made of self-contained statements that name their subject, make one specific claim and can be checked. The test is whether a key sentence still reads as complete and true when lifted out of the page with nothing around it.
The foundations overlap: clear structure and useful answers help both. The difference is the unit. Google ranks pages; assistants lift passages. So for AI search, the answer has to sit directly under a question-shaped heading, in a sentence that stands on its own.
Put the core answer in the first forty words under the heading, in one or two sentences. Detail, caveats and examples can follow. The extractable part is short; the page does not have to be.
No. Assistants look for the clearest available statement of a true thing, not keyword density. Use the words a buyer would actually type in your headings, then write plainly underneath.
Pricing, then the one-sentence description of what you do, then comparison pages. Those answer the questions buyers ask when choosing between vendors, which is where being quoted matters most.
No. Quotable content is necessary but not sufficient: the page must also be fetchable by the right agents and your company must be identifiable as an entity. And assistant output varies between runs, so nobody can guarantee a citation.