AI doesn't read your page — it lifts a chunk. Write for that.
You have a page that ranks, says the right things, and never gets cited by AI. The usual diagnosis — more authority, more links — misses the real problem. The model never read your page as a page. It chopped it into passages and quoted someone whose passage stood on its own.
the short answer
AI engines do not read your page top to bottom. They break it into passages, score each one on its own, and cite the ones that stand alone and answer directly. A 2025 study found 40-75 word answers cited about 3x more than longer ones. So lead every section with the answer to its implied question, keep passages self-contained, and use question-style headings, one-idea paragraphs and lists where they fit. The fix for "ranks but never cited" is usually structural, not authority.
key takeaways
- LLMs do not read your page top to bottom: they chunk it into passages, score each one independently, and cite the ones that stand alone and answer directly.
- A 2025 analysis of 10,000 citations found 40-75 word passages cited about 3x more than longer ones: the short, self-contained answer wins.
- Ranking is not enough — only ~12% of ChatGPT citations match Google’s first page; the reason you are not cited is usually structural, not authority.
- The number-one pattern is answer-first: resolve the section’s implied question in the first one or two sentences, before any elaboration or preamble.
- Patterns that help: question-style H2s, self-contained sections, one-idea paragraphs, fact density with named sources, and lists or tables where the content fits.
how often a passage gets cited, by length — 10,000-citation analysis, 2025
In words, so the bars do not carry it alone: an analysis of 10,000 AI citations in 2025 found a clear sweet spot in answer length. Passages of about 40 to 75 words were cited roughly three times as often as passages over 75 words, and meaningfully more than very short ones. The reason is mechanical, not stylistic: a 40-to-75 word passage is a complete answer a model can lift and quote with confidence — long enough to stand on its own, short enough that the system does not have to decide which part to trim. Write your lead answers to that length and you are handing the model exactly the unit it wants to cite.
Why structure beats authority here
The instinct when a page is not cited is to reach for the old levers — more backlinks, more domain authority, more words. For AI citation, those mostly miss, because the engine is not ranking your page against others; it is scanning it for a passage it can lift. A page can be authoritative, comprehensive and well-ranked, and still offer the model nothing clean to quote, because every answer is wrapped in three sentences of context and dependent on something said two sections earlier. Authority gets you considered; structure gets you quoted.
This is why the same brands that won classic search can lose AI citation to smaller, scrappier sites. The winners are not necessarily more authoritative; they are more extractable. They answer the question in the first line, they break a topic into self-contained sections, they state facts with specific numbers and named sources. None of that requires a bigger backlink profile — it requires writing the page so a machine that reads in passages can find a passage worth citing. That is a craft problem, not a budget problem, which is the good news: it is fixable by anyone willing to restructure what they already have.
The chunk rule, in three parts
How AI actually reads a page, the answer-first pattern that follows from it, and why ranking does not save a page that buries its answers. Open each layer for the part that changes how you write.
01 How AI actually reads a page
The mental model to drop is that a model reads your article the way a person does, start to finish, forming an impression. It does not. A retrieval system breaks your page into passages, scores each passage on its own for how well it answers the query and how confidently it can be quoted, and assembles an answer from the winners — attaching citations to the sources it lifted. The consequence is blunt: a brilliant insight buried in the third sentence of a dense paragraph, dependent on setup from two sections earlier, gets passed over, not because it is wrong but because extracting it cleanly is too risky for a system that needs to be sure it is quoting you accurately. The unit of citation is the passage, not the page, and most pages were written for the page.
02 The answer-first pattern
The single highest-value change is to lead every section with the answer. Take the question a heading implies and resolve it in the first one or two sentences, before any context, history, or "in this section we will" throat-clearing. The data is specific: passages of roughly 40 to 75 words are cited far more than longer ones, because that is a complete, liftable answer a system can quote without trimming. The structure is simple — first sentence answers the heading question directly, the next one or two add essential qualifiers, everything else moves down. What you are removing is the narrative warm-up that pushes the answer out of the prime extraction window. Readers benefit from the same move; nobody resents getting the point first.
03 Why ranking is not enough
It is tempting to assume that if you rank, you will be cited, and the data says otherwise. Independent 2025 research found only around 12% of the URLs cited by ChatGPT and its peers rank in Google’s top ten for the same query, and most do not appear in the top hundred at all. AI is not drawing from the same pool that twenty years of SEO optimised for. Domain authority still helps, but it does not rescue a page that buries its answers, lacks clear question-and-answer structure, and reads as one long narrative. A weaker site that is easy to extract from will out-cite a stronger one that is not. That is the uncomfortable, liberating part: the gap is usually structural, which means it is fixable on the page you already have.
The patterns that get pulled
Beyond the lead answer, a handful of structural habits make a page reliably more extractable, and they reinforce each other. Use H2 headings that match the natural-language question a reader would actually ask, so the model can map a query straight onto your section. Make each section self-contained, so a passage lifted from the middle of the page still reads as a complete thought without the setup above it. Keep paragraphs to one idea each, because a paragraph carrying five claims is hard to chunk and easy to skip. And raise the fact density: specific numbers, dates, named entities and attributed sources give a retrieval system the verification signals it needs to quote you with confidence, where vague generality gives it nothing to hold.
Format is part of the structure, not decoration on top of it. Where your content is genuinely a process, a numbered list creates clean extraction boundaries the model reads as discrete steps; where it is a comparison, a table makes the relationships explicit in a way flowing prose cannot. Research on retrieval has found structured elements like tables improve extraction accuracy, because they impose the semantic boundaries that line up with how systems chunk content. The discipline is to match the format to the shape of the information — list what is a list, tabulate what is a comparison, and write clean answer-first prose for everything that is an argument or an explanation. Forcing everything into one format is its own way to lose.
What not to do in AI's name
The failure mode to avoid is over-optimisation — turning "structure for AI" into robotic, keyword-stuffed, list-everything pages that read as though no human was meant to see them. It does not even work: the Princeton GEO research that found structured, fact-dense content lifting AI visibility by thirty to forty percent also found that keyword stuffing decreased visibility. Models are not fooled by the surface tricks that once moved rankings; they are scoring whether a passage genuinely, cleanly answers a question, and a page engineered to look optimised rather than to be useful scores worse, not better. The honest target is clarity, and the patterns above are simply what clarity looks like to a system that reads in passages.
It helps to remember who you are still writing for. The patterns that make content extractable for a model are the same ones that make it readable for a person who skims, looks for the point, and leaves if they cannot find it. There is no real trade-off between writing for AI and writing for humans, only between writing clearly and writing badly.
Writing for the chunk: quick answers
Does writing for AI mean writing in fragments?
No, and that is the most common misreading. Writing for extraction means making each section answer its own question clearly and early, not chopping prose into disconnected fragments. A self-contained chunk is still a real paragraph — it just leads with the answer instead of burying it, and does not depend on context three sections up to make sense. Good writers have always done a version of this: state the point, then support it. What changes for AI is the discipline of doing it in every section, so that any passage a model lifts still reads as a complete thought. The result is usually clearer for humans too, because readers also skim and also resent hunting for the point. Fragmentation makes content worse for everyone; structure makes it better for both audiences at once.
How long should an answer paragraph be?
Short enough to lift cleanly, long enough to stand alone — and the data points at a specific window. A 2025 analysis of 10,000 AI citations found that passages of roughly 40 to 75 words were cited far more often than longer ones, and meaningfully more than very short ones. The practical target is a lead answer of two to four sentences: the first directly resolves the question implied by the heading, the next one or two add the essential qualifying context, and anything beyond that moves to a follow-up paragraph. Too long and the model has to decide which part to quote; too short and the passage lacks the context to stand on its own. The 40-to-75-word lead is the sweet spot because it is a complete answer a system can quote with confidence and without trimming.
Do lists and tables really get cited more?
They tend to, for a mechanical reason rather than a stylistic one. Numbered lists and tables impose explicit boundaries and relationships that line up with how retrieval systems chunk and parse content, so the model has less interpretation work to do and more confidence about what each element means. A process is clearer as numbered steps than as a flowing paragraph; a comparison is clearer as a table than as prose weaving four options together. That said, the format has to fit the content — forcing everything into lists is its own failure, producing thin, choppy pages. The honest version is: where your content is genuinely a sequence, a comparison, or a set of discrete items, structuring it as such makes it more extractable; where it is an argument or an explanation, clear answer-first prose does the job. Match the format to the shape of the information.
Will structuring for AI hurt my human readers?
Done well, it helps them — done badly, it hurts both. The patterns that make content extractable for AI are mostly the patterns that make it readable for people: a clear answer up front, descriptive headings, one idea per paragraph, concrete facts over vague claims. Readers skim, look for the point, and leave if they cannot find it, so leading with the answer serves them as much as it serves a model. Where it goes wrong is when teams over-optimise into robotic, keyword-stuffed, list-everything pages that read as though written for a machine — which, research found, actually decreased AI visibility too. The failure mode is not "writing for AI"; it is writing badly in AI’s name. Write clearly, lead with the answer, and you serve both audiences without choosing between them.
A note on sources and certainty
The figures here — the 40-to-75-word sweet spot from a 10,000-citation analysis, the roughly 12% overlap between AI citations and Google’s first page, the Princeton finding that structure lifts visibility while keyword stuffing lowers it — come from 2025 research into how generative engines pull sources. We present the exact numbers as a snapshot, because retrieval systems retune and the precise figures will move; what is durable is the mechanism, which is unlikely to reverse: models read in passages, and passages that stand alone and answer directly get cited. We have kept to what was measurable as of this writing. The AC Group has spent 27 years arguing that being useful beats gaming the system, and this is that argument made concrete at the level of the paragraph: write the answer cleanly, and the machine has something worth quoting.