How do I write an article that AI cites?

What makes an article citable?

An article is citable when a single passage can be lifted out of it and dropped into an answer without losing meaning. That is the whole difference from classic writing. A human reads top to bottom and builds up context along the way. A language model searches many documents for the spots that match the question asked and formulates an answer from them. The link in the citation points at the whole page, but what gets selected and processed is the individual passage. So what is citable is not the article, it is the paragraph.

From that follows one rule that governs everything else: every paragraph has to stand on its own. No back-references like “as mentioned above”, no pronouns whose subject sits two paragraphs back, no statement that falls apart without the heading above it. If your best sentence only makes sense after someone has read the three paragraphs before it, it is worthless as a quote.

How do you build the answer block at the top?

The answer block is the most important paragraph of the article: two to four sentences at the very top that answer the title question completely. Completely means that someone who reads only this block has a correct answer and needs the rest only for depth. Three things belong in it:

• The term from the question, spelled out: if the heading reads “What is llms.txt?”, the answer starts with “llms.txt is ...”, not with “It is a file that ...”. The model needs the anchor inside the sentence, not just in the line above it.
• The answer without qualifiers in the first sentence: conditions, exceptions and special cases come in sentence two or three.
• The distinction, if the topic is often confused with something else: one sentence is enough, namely what it is not.

What does not belong in it is preamble. Sentences like “A lot has changed in recent years” cost you the passage, because a system scans the opening lines and finds nothing usable.

How do you structure the rest of the article?

After the answer block, every section becomes its own small answer unit. Phrase every H2 as a question, the way your audience would ask it, and answer it in the first sentence below. The rest of the section delivers reasoning, example and caveat, in that order.

Why questions? Because a query in a chat interface is a question itself. The closer your heading sits to the phrasing someone types, the clearer the match between question and passage. The same holds for people: a question as a heading promises an answer, a noun promises nothing. Four measurements that have proven themselves in practice:

• One question per H2, not two: double questions produce sections that are only half citable.
• Paragraphs of three to five sentences: shorter reads as chopped up, longer mixes two thoughts into one passage.
• Lists only for items of equal rank: anything that explains a connection belongs in running text. A list of sentence fragments is hard to cite, because each item lacks context.
• One FAQ block at the end: for the questions that belong to the topic but would break the line of argument.

How do you back up claims so an AI will use them?

A number without a source and a date is a risk for an answer system: it can't trace where the value came from, and when in doubt it reaches for a source where it can. So the originator and the time period belong right next to the number, in the same sentence, not in a footnote at the bottom of the page.

Concretely: “Google's share of verified bot traffic stood at 30.1 percent in the second quarter of 2026 (Cloudflare Radar)” is citable, as shown in Who is reading your website? “Google's share is falling sharply” is not. The same goes for terminology: whoever supplies a definition that others can lift word for word gets cited as the definition.

Two points that often get lost here. Link the source inside the text, not only in a list at the bottom, so the attribution sits inside the quoted paragraph itself. And keep the date of the last update visible, in the text and in the structured data. Freshness is a selection criterion in its own right for several answer engines, How do I get cited by Perplexity? covers that in detail.

Which mistakes cost you the citation?

The most common mistakes are not writing mistakes, they are structural ones:

• The run-up: three paragraphs of lead-in before the question is even asked. The citable part then sits halfway down the page.
• Marketing speak instead of statements: “the smartest solution on the market” contains nothing a model could use. Test every sentence with one question: could this even be disproven?
• Key content only in images: tables as graphics, numbers in the infographic, the core message in the video. What isn't on the page as text doesn't get cited.
• Content loaded via JavaScript: the common AI crawlers don't execute JavaScript and only see the server-rendered HTML. Google's systems do render, but you can only rely on that at Google. A look at the page source shows you what survives.
• One topic spread across five pages: if three pages half-answer the same question, they compete with each other. Better one page that closes the question.

How do you know what to write about?

The best article structure is worth nothing if nobody asks the question. The process before writing matters just as much as the writing: first find out which questions your audience actually asks, then check which of them your website already answers, and write only the gaps. How to find those questions: These tools find the questions users ask ChatGPT and Perplexity. Whether the answers name you afterwards: How do I measure my visibility in AI answers?

That is exactly the loop flize bundles: GEO score, question scouting, answer check and content production, meaning articles written to be cited, for your own domain or for clients.