An AI answer engine never reads your article. It breaks your page into passages, scores each passage on its own, retrieves the ones that look like they resolve a question, and assembles a reply from the winners. That one mechanical detail explains nearly everything about why some pages get quoted constantly and better pages get ignored: a 3,000-word masterpiece can lose to a single tidy paragraph on a smaller site, because the engine was never comparing pages.
Which leads to the rule that governs all of this: write for the chunk, not the page. Your unit of optimisation is no longer the article; it is the individual block that has to survive being ripped out of its surroundings and dropped into someone else's answer. Below are the craft-level rules for writing those blocks, and the trap to avoid, which is turning your whole site into beige reference text in the process.
This is a writing-craft companion to our plain-English guide to the AI search acronyms, which covers what AEO is and how it relates to GEO and LLMO. This piece assumes you know that and gets straight to the page.
What actually gets pulled
Passages that end up quoted in AI answers share a recognisable profile. They tend to run roughly 75 to 150 words: long enough to carry a complete thought, short enough to stay on one. They open by stating the answer rather than building toward it. They are dense with specifics, meaning named things, numbers and dates rather than adjectives. And crucially, they make complete sense with zero surrounding context.
That last property is not a style preference, it is a retrieval constraint. Systems typically divide documents into chunks of a few hundred words before scoring them, so a sentence whose meaning depends on the paragraph above it may be evaluated without that paragraph ever being present. Self-contained beats comprehensive, every time.
| Element | Hard to quote | Easy to quote |
|---|---|---|
| Opening | "It can significantly reduce costs." | "Server-side tracking reduces data loss from ad blockers." |
| Heading | "Our Approach" | "How long does server-side tracking take to set up?" |
| Claim | "Results improve considerably." | "Setup typically takes two to four weeks for one site." |
| Structure | Three sentences of context, then the point | The point, then three sentences of context |
Rule one: never start a block with a pronoun
This is the highest-return edit available, and almost nobody makes it. Paragraphs that open with "it", "this", "they", "these tools" or "as mentioned above" are unquotable by construction, because the thing being referred to lives in a chunk the retrieval system may never see. "It reduces data loss" means nothing on its own. "Server-side tracking reduces data loss" travels anywhere.
The fix costs nothing but feels slightly wrong: name the subject again, even when it reads as mildly repetitive top-to-bottom. That small redundancy is the price of independence, and readers barely notice it because they are usually skimming to the part they need anyway. Apply the same logic to sections that open with "Another benefit is…", which silently depends on a list the engine cannot see.
MagicPost
AI LinkedIn post generator with scheduling, analytics and a large inspiration library for creators.
Best for: Professionals & creators growing their LinkedIn presence
Take any block on your page, cover everything above and below it, and ask four questions:
→ Does it name its subject, or does it rely on a pronoun pointing somewhere else?
→ Does it answer something a person would actually type or say?
→ Does it contain a specific — a number, a date, a named tool, a defined range?
→ Would it be true and complete if pasted into a stranger's document?
If any answer is no, that block cannot be cited. Rewriting it usually takes under two minutes, and the fix is almost always the same: move the answer to the front and replace the pronoun with the noun.
Rule two: answer first, expand second
Put the answer in the first sentence under the heading, then use the following 40 to 70 words to complete it cleanly. Everything else — nuance, caveats, examples, the interesting digression — goes underneath. Long wind-ups are the single most common reason good content never gets quoted, because the engine finds three sentences of scene-setting where it needed a definition.
The heading matters just as much, and should be phrased the way a person asks the question rather than the way your team names things internally. "Pricing Structure" tells a retrieval system nothing; "How much does an affiliate program cost to run?" matches an actual query almost word for word. This is the same instinct behind classic on-page SEO, sharpened: you are matching phrasing to a spoken question, not to a keyword with search volume.
The rule An AI never reads your page. It reads pieces of it, scored on their own. So write for the chunk, not the page: every block should make complete sense to someone who will never see the paragraph above it.
Rule three: be specific enough to be worth quoting
Models reach for passages carrying their own evidence. A sentence containing a number, a date, a named product or a defined range is more useful to an answer engine than a sentence full of confident adjectives, because the specific one can be reproduced without the engine having to vouch for a vague claim. "Costs drop substantially" is unusable. "Costs typically fall by 20 to 30% in the first quarter" is quotable, attributable and checkable.
Dates matter for the same reason. Answer engines favour content that signals recency, so stating when something was measured or last reviewed makes a passage safer to reuse. This is also where original material earns its keep: your own data, benchmarks and first-hand results are simultaneously the most quotable and the least replaceable things you publish, which is the deeper argument in our GEO playbook and in building a content strategy that compounds.
Rule four: one idea per block
Keep paragraphs to two to four sentences and give each section a single job. A block covering three loosely related points dilutes its own signal: retrieval works on how strongly a passage matches one question, so a passage answering three questions half-way matches none of them well. Write core facts as clean subject-verb-object statements, because those parse reliably in isolation.
Structure the page so the boundaries fall where you want them. Lists and comparison tables are extraction-friendly formats, since each row or item is already a discrete unit. Aim for sections of roughly 100 to 300 words so a natural chunk boundary lands between your ideas rather than through the middle of one, and keep the visual structure clean for the same reason we described in designing websites for AI search. None of it works, of course, if the page cannot be crawled and rendered in the first place, which is where technical SEO fundamentals remain the gate.
The trap: quotable is not the same as beige
Follow the advice above too literally and you produce a site that reads like an encyclopaedia written by a committee: correct, extractable, and completely forgettable. That is a real cost, and it is worth naming, because the same brand that gets quoted also has to convince a human once they arrive.
The way out is to split the labour within each section. The first 40 to 70 words under a heading are the answer: plain, direct, specific, built to be lifted. Everything beneath is yours, where judgement, opinion, examples and voice belong. There is a commercial argument too. If your page says precisely what every competitor says, an AI can satisfy the question without naming anyone in particular; distinctive data and a genuine point of view are what make you the source worth attributing. Being summarised into invisibility is the real risk of sameness, and it is the same dynamic behind the metric shift in zero-click content.
What not to bother with
Two habits waste time. The first is bolting a long FAQ section onto every page purely for the format, filled with questions nobody asks, which adds length without adding a single retrievable answer. Write FAQs only for real questions, drawn from what buyers actually ask your sales team or search for, and answer each in one self-contained block. Getting those questions in front of people still depends on distribution, since an unread page is never retrieved either. The second is stuffing question-shaped headings with keywords until they stop sounding like something a person would say, which defeats the purpose of matching conversational phrasing.
It is also worth being realistic about results. Getting quoted is volatile, the sources cited for a given question change frequently, and no formatting rule guarantees inclusion. Treat AEO as improving your odds across many questions rather than as a switch, keep an owned channel such as a newsletter as the steady counterweight, and measure presence over time rather than judging a week, which is the same caution we raised about AI Overviews and about local AI discovery.
The short version
AI answer engines never read your page; they split it into passages, score each one independently, and quote the passage that resolves the question, which is why a long excellent article can lose to one tidy paragraph. Write for the chunk, not the page. Passages that get quoted run roughly 75 to 150 words, open with the answer, carry specifics, and make complete sense with no surrounding context. Four craft rules deliver that: never start a block with a pronoun, because retrieval may never see what "it" refers to; answer first and expand second, under a heading phrased the way people actually ask; include a number, date or named thing so the passage carries its own evidence; and keep one idea per block in sections of about 100 to 300 words so chunk boundaries fall between your ideas. Avoid the trap of writing everything this way, since a site of pure reference text has no reason to be cited over anyone else's: keep the answer clean and put your judgement, data and voice in the expansion beneath it.
Want content that AI tools actually quote?
We help brands write pages built to be cited, without losing their voice.
Explore Content Marketing →Frequently asked questions
What is answer engine optimization?
Answer engine optimization, or AEO, is the practice of writing and structuring content so AI systems can lift a passage from it and quote it as the direct answer to a question. Where classic SEO competes for a position in a list of links, AEO competes for a sentence inside a generated answer in tools such as ChatGPT, Perplexity, Gemini and Google's AI features. The mechanical reason it differs from ordinary writing is that answer engines do not read pages the way people do. They break documents into passages, retrieve the passages that seem to resolve a question, and assemble an answer from them. That means a long, excellent article can lose to a single well-formed paragraph elsewhere, because the engine is not comparing page quality, it is choosing the chunk that answers the question cleanly. AEO sits on top of SEO rather than replacing it, since a page still has to be crawlable and credible before any passage in it can be retrieved.
How long should an answer block be for AI to quote it?
Aim for a direct answer of roughly 40 to 70 words immediately beneath a question-shaped heading, sitting inside a section of about 100 to 300 words. Analysis of passages that get pulled into AI answers commonly finds them in the range of about 75 to 150 words, which is long enough to carry a complete thought and short enough to stay focused on one. The surrounding section matters too, because retrieval systems typically split documents into chunks of a few hundred words, so a section of that size tends to survive as one coherent unit rather than being cut in half. The practical rule is one idea per block: state the answer, then expand underneath with nuance, examples and caveats. Long wind-ups are what kill extractability, since a paragraph that spends three sentences setting up context before reaching the point gives the engine nothing clean to lift.
Why do pronouns hurt AI citations?
Pronouns at the start of a paragraph break the passage's independence, and retrieval systems score passages independently of the text around them. A paragraph beginning 'It reduces costs significantly' is meaningless on its own, because whatever 'it' refers to lives in a previous paragraph the retrieval scorer may never see. The same applies to 'this approach', 'they', 'these tools' and 'as mentioned above'. Documents are typically divided into chunks of a few hundred words, so any sentence that depends on earlier text is at risk of landing on the wrong side of a boundary. The fix is simple and costs nothing: name the subject again. 'Server-side tracking reduces data loss from ad blockers' is quotable anywhere; 'It reduces data loss' is quotable nowhere. Repeating nouns feels slightly redundant when you read the page top to bottom, which is exactly why writers avoid it, but that small redundancy is what makes each block able to stand alone.
Does writing for AI make content sound generic?
It does if you apply the rules to every sentence, but that is a choice rather than a requirement. The structural advice behind AEO, meaning direct answers, self-contained blocks and plain subject-verb-object sentences, applies best to the answer itself, not to the entire page. A useful split is to keep the first 40 to 70 words under each heading clean and quotable, then use the expansion beneath it for the judgement, examples, opinions and voice that make the piece worth reading once a human arrives. There is also a commercial argument for keeping the distinctive parts: if your content says exactly what every competitor says, an AI can summarise the topic without needing you, and there is no reason to cite you specifically. Original data, first-hand experience, a clear point of view and specific numbers are simultaneously the most quotable and the least replaceable material you can publish, so the goal is precision in the answer and personality around it.
