Answer first, qualify second: writing to be picked up by an AI

For twenty years we wrote for a reader who arrived through the home page or a search result, then read from top to bottom. Generative engines read differently. They break a question into sub-questions, look for the passage that best answers each one, and assemble a reply from fragments taken across several pages. Google describes that mechanism as query fan-out in its AI Mode announcement. Writing to be picked up means writing in self-contained passages, each one opening with its answer.

The unit of selection is the passage

When a model prepares an answer, it does not read your page as a whole. It receives extracts, often a few hundred words, selected because they match a sub-question. An extract that opens with "as we saw above" or "this point deserves some explanation" is unusable on its own: it depends on what came before. An extract that opens with "the statutory cooling-off period is fourteen days from delivery" is usable immediately.

Every section therefore has to read without the others. Its heading poses an identifiable question, its first sentence answers it, and the rest justifies, qualifies and points to sources. Human readers gain from this too, because it is exactly how they scan a long article.

From question to sourcing: the order in which a section gets written

1. One question per H2. The section heading is a real question, phrased as a reader would ask it, or its declarative equivalent. "How much does an SEO audit cost?" beats "Pricing". The sub-questions an engine generates look like reader questions, not like outline headings.
2. The answer in the first sentence. No preamble, no restatement of context. The figure, the definition, the position. If the answer depends on the case, give the most common case first, then flag the exceptions.
3. The qualification next, with its source. Conditions, limits and counter-examples come after the answer, and every claim that matters points to an identifiable source. Research on optimizing for generative engines, in particular the GEO study published in 2023, measures that adding references, statistics and attributed quotations is what raises the odds of a passage being picked up the most.
4. Close on what the reader has to check. The last sentence of the section says what the reader should do or verify, or hands over to the next section. Wrap-up formulas give nothing to the engine or to the reader.

Before and after, on the same paragraph

An example taken from a guide on sitemaps shows the difference.

The second version answers, gives a figure, points implicitly to official documentation, and stands alone. The first says nothing until you have read what follows. An engine that has only this fragment in front of it will take the second.

What Google says about its own systems

Google's documentation on AI features states something precise: there is no special markup for being cited in an AI Overview, and the recommendations are those for helpful content. The guide on creating helpful content asks questions that overlap with the method above. Does the content bring original information, analysis that goes beyond the obvious, a complete answer to the question asked? Is it written by someone who clearly knows the subject? These are editorial questions, not technical ones.

Bing publishes webmaster guidelines that point the same way and note that its generated answers draw on the classic search index: what is not indexed cannot be cited.

The formats that help, and the ones that hurt

Some structures make a passage easier to extract and to reproduce:

• The comparison table for any question of the "what is the difference between A and B" kind. A model reproduces it faithfully, usually with attribution.
• The numbered list for a procedure, provided each step is a complete, self-contained sentence.
• The one-sentence definition at the start of a section, where the subject allows it.
• The date and the unit on every figure: "14 days", "in 2026", "in US dollars, before tax". A figure with no unit and no date is rarely picked up.

The article title obeys the same rule. An H1 that carries the reader's real question, in their words, gets found more often than a catchy but vague headline. "Do you have to label your articles as AI-generated?" matches a user's question; "AI and transparency: a new challenge" does not. The same goes for the meta description, which is not a slogan but the condensed answer the engine will display or reuse.

Three habits inherited from volume SEO degrade pickup: repeating the keyword in every paragraph, which the GEO study measures as ineffective and sometimes counterproductive; transition paragraphs that say nothing; and introductions that announce the outline instead of getting into the subject.

Holding the method across three hundred articles

This method is easy to apply to one article and hard to hold across three hundred. It assumes the writing template enforces the question, answer, qualification, source structure, that sourcing is verified before publication, and that review covers how quotable the passages are as much as the spelling. That is where the production tool makes the difference: a CMS that generates articles without enforcing this discipline produces pages nobody cites. Another article covers the four signals generative engines look at, of which the one described here is only the second.

One neglected point: writing in self-contained passages also protects you against being quoted out of context. If each section carries its own qualification, a model that quotes it takes the qualification with it. If the qualification sits three sections away, it gets lost, and it is your name attached to a truncated claim.