Somewhere in your best article is the sentence that answers the question. It is sitting in paragraph six, after the context and the setup, exactly where a trained writer puts the payoff.
An AI engine will read it and quote your competitor anyway.
That is not a verdict on your writing. Citation engines do not read articles the way people do. They pull passages, score them one at a time, and pick whichever one answers the question without needing the rest of the page. Kevin Indig’s analysis of 1.2 million AI answers and 18,012 verified citations found that 44.2% of ChatGPT’s citations came from the first 30% of a page. The middle stretch took 31.1%. The last third took 24.7%, falling off hard near the footer. 1 Source 1 Kevin Indig, “The science of how AI pays attention,” Growth Memo, February 16, 2026. Analysis of 1.2 million AI answers and 18,012 verified citations. Placement: 44.2% from the first 30%, 31.1% from the middle, 24.7% from the final third.
That curve is not really about position. The top of a page is where most writers state things plainly. Everywhere else, they build.
How LLMs Decide Which Sources to Cite covers the retrieval machinery underneath all of this. What follows is the writing side: four habits that decide whether a passage can be quoted at all, none of which require buying anything.
What does answer-first writing actually mean?
Answer-first writing puts the direct answer in the opening sentence of a section and spends the rest of the paragraph earning it. Evidence follows the claim instead of leading to it. Nothing gets saved for the conclusion, because the conclusion is the part nobody retrieves.
Most business writing runs the other way. The heading names a topic. The first sentence sets up context. By the third or fourth sentence you finally say the thing, and by then a re-ranker scanning for a clean passage has already scored your paragraph against someone else’s and moved on.
It helps to stop thinking of your page as a document. To a citation engine it is a bag of passages, each one competing on its own merits against passages from every other site. A 4,000-word guide with nothing quotable in it loses to a thin post with one sharp paragraph.
The evidence that structure alone moves citations
Structure moves citations independently of what you actually say. Four researchers from the University of Tokyo, Tsukuba, Hiroshima and Japan’s National Institute of Informatics took 200 articles, produced a restructured version of each, and ran both through six generative engines. Citation rates rose 17.3%. 2 Source 2 Yu, Yang, Ding, Sato. “Structural Feature Engineering for Generative Engine Optimization: How Content Structure Shapes Citation Behavior.” arXiv preprint 2603.29979, March 31, 2026. 200 articles in baseline and restructured versions across Google SGE, Bing Chat, Perplexity, Phind, ChatGPT and Claude, for 2,400 test cases. Citation rate +17.3%, subjective quality +18.5%. Semantic preservation verified by embedding similarity, mean 0.843. Preprint, not yet peer reviewed.
The design is the interesting part. They were not comparing good content to bad content. They restructured the same articles and confirmed the meaning survived the edit by checking embedding similarity, which landed at 0.843. Same substance, rearranged, and the engines quoted it noticeably more. The restructured versions also scored 18.5% higher on subjective quality, so this is not a case of writing worse to please a machine.
It is a preprint rather than peer-reviewed work, and 200 articles is a modest sample. Take the direction seriously and hold the decimal point loosely.
Agency-side numbers point the same way. A European GEO shop reported that across 27 client sprints, pages rewritten answer-first with sourced statistics and FAQ schema averaged 4.2 AI citations a month on their target queries, against 0.8 for pages left in narrative form. 3 Source 3 PingPrime, “Optimisation pour les moteurs IA,” updated July 24, 2026. 27 GEO sprints across Belgium and France, 2025 to 2026. Answer-first pages with sourced statistics and FAQPage schema averaged 4.2 AI citations per month against 0.8 for narrative pages. Vendor self-report. That is a vendor reporting on its own work, so treat it as a direction rather than a benchmark. It is worth noting anyway because it isolates the same combination the research points to.
Put the answer in the first sentence
The habit is simple to describe and genuinely hard to practice: every section opens with its claim, and every paragraph inside it opens with its own smaller claim. Paragraph three should never read as though paragraph one is required reading. Assume the reader landed cold, because a machine did.
Almost everything in a writer’s training pushes against this. Academic writing builds toward a thesis. Journalism buries the lede on purpose to create tension. Both traditions treat the answer as a reward for staying, which was a reasonable bet when the only reader was a human who had already decided to stay.
The answer is not a reward for reading. It is the first sentence, and everything after it is evidence.
There is an editing pass that catches this in about two minutes. Read only the first sentence of every paragraph, in order, and ignore the rest. If those sentences alone tell the story of the section, the structure is right. If they read like a list of topics you are about to cover, you have written setup and called it content.
Do question headings actually matter?
Question headings matter where the section genuinely answers a question someone would type. Indig’s data found cited content was twice as likely to contain a question mark, and 78.4% of citations tied to a question came from headings rather than body copy. 4 Source 4 Indig, same study. Cited content was twice as likely to include a question mark, and 78.4% of citations tied to a question came from headings. Engines appear to read an H2 as a prompt and the paragraph beneath it as the candidate answer.
Which is a useful mental model for the whole exercise. Your H2 is the question. The 40 or so words under it are your submitted answer. Every heading in your page is a small bid, and half of them are currently bidding with topic labels.
The counsel to make every heading a question is where this tips into parody, though. A page of nothing but questions reads like a support portal and gives a skimming human no map of your argument. Use questions where the section answers one. Use a declarative heading where the section makes a point. The test is whether a buyer would type your heading into a search box, and plenty of good sections will honestly fail that test.
Write stats that survive being quoted
Write every statistic so it holds up alone, which takes three parts in one sentence: the figure, the source, and the scope. A number that depends on the sentence above it will never be lifted, because the engine lifting it has no access to the sentence above it.
“The number nearly doubled” means nothing on its own. “Citation rates rose 17.3% across six generative engines when 200 articles were restructured without changing their substance (arXiv preprint, March 2026)” survives anywhere you drop it. The reader knows what moved, by how much, in what conditions, and who says so.
This is also the discipline that keeps you honest, which is the part nobody mentions. Forcing the scope into the sentence exposes the claims you cannot actually support. Every stat that resists the format is usually a stat you half-remembered from a vendor blog that was citing a second vendor blog.
FAQ blocks work the same way one level up. Each answer should read as a standalone paragraph: the question, the direct answer, one concrete number, the implication. The visible answers are what engines quote. FAQPage JSON-LD 5 Definition 5 FAQPage schema: a JSON-LD block marking up question and answer pairs so crawlers can extract them without parsing prose. Google stopped showing FAQ rich results in May 2026. can mirror them at no cost if your templates generate it, which is a platform question AEO Is an Architecture Problem takes up.
Say it plainly or lose the passage
Define things directly. Indig found cited passages were nearly twice as likely to use plain definitional constructions, the “X is” and “X refers to” phrasings that most editors trim as clunky. 6 Source 6 Indig, same study. Cited passages were nearly twice as likely to use clear definitional constructions such as “X is” and “X refers to.” The study measured definitional phrasing, not hedged language. The clunky version is the quotable version, because it hands the engine a complete statement instead of an implication.
Hedging is the same problem wearing different clothes. Nobody has measured “may” and “might” against citation rates, so treat this as craft rather than data: a passage stuffed with qualifiers gives a re-ranker less to work with than one that commits. “Retrieval models lift clean passages” and “retrieval models may tend to lift cleaner passages” make the same claim, and only one of them is worth quoting.
None of which is a license to overstate. Precision is what makes a sentence citable, and precision sometimes means saying that a sample was small or that a finding is a vendor’s own. “A March 2026 preprint measured a 17.3% lift across six engines” is confident and accurate. “Better structure will increase your citations” is neither.
One paragraph, rewritten
Here is the pattern in its natural habitat. First, the version most teams publish:
There are several factors that influence whether AI engines choose to cite a given piece of content. Some research suggests that structure may play a role, and practitioners have noted that pages with clear organization tend to perform better. The specific techniques driving this effect are still being studied, but answer-first approaches appear to be gaining traction.
Four sentences, five hedges, no claim. Now the same paragraph written answer-first:
Restructuring content lifts AI citation rates without changing what the content says. A March 2026 arXiv preprint restructured 200 articles and measured a 17.3% citation increase across six generative engines, with meaning held intact. The techniques are answers in the opening sentence, question headings where they fit, self-contained statistics, and plain definitional phrasing.
Same topic, same length, and the second one can be quoted by anything that finds it. The first cannot be quoted at all, because there is nothing in it to quote.
Why we write this way
We spent 16 years writing the first version, at scale, for a living.
Inbound training taught a specific shape. Open with context so the reader settles in. Build the argument across the whole page so they stay on it. Land the payoff late, where it rewards the people who made it that far. That shape was tuned for a ranked blue link and a human who had already committed to reading, and it worked for about a decade.
The habit outlived the conditions. When we rebuilt our own library, the expensive part was not learning the new rules, which fit on an index card. It was unlearning the instinct to warm up. Our drafts kept arriving with two throat-clearing sentences in front of every good one, written by people who knew better and did it anyway, because that is what sixteen years of muscle memory feels like.
So we edit for it explicitly now. First sentence of every paragraph, read in sequence, on every draft that goes out. It catches the throat-clearing faster than any tool we have tried, and it costs one pass.
What you actually get
The payoff is not really more citations, though that follows. It is that you stop needing permission to fix your own content.
Most of AEO sits behind somebody else’s approval. Rendering is a developer sprint. Schema coverage is a platform decision. Crawler access lives in a config file that marketing does not own. The four layers all matter, and three of them require a budget conversation before anything moves.
This one does not. It is a habit your existing team can adopt on Monday, applied to pages you have already paid to produce, with no migration and nothing to procure. That makes it the first thing to fix, not because it is the most powerful lever, but because it is the only one you can pull by yourself.
Your page passes if all of this is true:
- Every section that answers a question opens with a direct sentence and lands the whole answer inside 40 to 60 words
- Several headings are phrased as questions a buyer would actually type
- Every statistic carries its figure, its source, and its scope inside a single sentence
- Terms are defined with plain “X is” constructions rather than described sideways
- Hedges appear only where the underlying claim is genuinely uncertain
- The FAQ block runs at least three items, each self-contained and under 60 words
- Reading only the first sentence of each paragraph still tells the story
Structure only pays off if the crawler gets to the page. The Free AI Website Assessment checks whether yours is readable at all.
Start with the pages that already earn traffic. Read each section opening as though you arrived there from nowhere, which is how the engine arrives. Cut the warm-up sentence. Move the claim to the front.
The work is unglamorous and it does not require anyone’s approval, which is the best argument for doing it first. Everything else in the AEO stack costs money. This costs a second pass.
The research is free. So is the diagnosis. 60 seconds →