How to write content that AI cites
A page can be well written, well designed, well ranked and still never get quoted. What decides it is whether the answer is near the top, written as text, and about one question rather than all of them.
AI sessions measured across the sites we track, last 30 days.
Quick answer
Content that AI cites answers one question, near the top, in text a model can lift without editing. We measured it with 29 of the 57 questions we track, all of them about one cenote park in Punta Cana, across 5 AI surfaces, which is 145 of the 285 measurements. The answers were built from 784 citations across 271 domains, and the measured site supplied 73 of them. Being one of the 271 is the whole job, and it is decided page by page.
The two failures that keep a good page out
The first is the buried answer. A page opens with three paragraphs of atmosphere and puts the price, the opening hours and the location halfway down. A human scrolls. A model that is assembling an answer takes what is near the top and moves on.
The second is the fact that only exists in a picture or a widget. Rates inside the booking engine, opening hours in a designed image, the distance to the airport in a map graphic. None of it is text, so none of it can be quoted, and the assistant fills the gap with a third party page that wrote the same fact in plain words.
Both failures pass every visual review, because the page looks finished. They only show up when you measure whether anything on it got quoted.
Not every surface reads the same way
The measured split is the argument against writing for one assistant. Out of those 145 readings, the site was named inside the answer text by Google AI Mode 14 times and by ChatGPT 9 times. Claude, Gemini and Perplexity named it zero times. Yet Perplexity linked to it 20 times, Claude 16 and Gemini 15, all three more often than ChatGPT, which linked it 9 times.
Read that carefully and the instruction is not to pick a surface. It is to write a page that works in both modes at once: quotable in a sentence for the surfaces that write names, and complete enough to be worth listing as a source for the surfaces that mostly link.
The practical version is a page that opens with a two sentence answer containing the name and the number, then earns the citation with the detail underneath. And read the split rather than the score: those 7 domains are our own travel websites, not hotel clients, and one pass is one reading, which proves nothing on its own.
One question, one page
The page that answers everything answers nothing quotably. A model asked how far a property is from the airport will pick the page whose subject is that distance over the page where the distance is sentence forty of a general overview.
So the unit of work is a question a guest actually types, and the sequence is boring on purpose: collect the questions, check which ones the assistants currently answer with somebody else's page, write the missing ones, then rerun the pass and see whether your URL entered the source list. The content plan in reporte.ai builds that queue from the prompts where you are absent rather than from a keyword tool, and the free AI Readiness Check scores whether the pages you already have are readable at all.
FAQ
A direct answer in the first lines, facts written as text, and one page per question instead of one page for everything. The models we track do not lift from images or from a booking widget, and in the corrected window they cited 784 sources across 271 domains, so a page that buries its answer is competing badly for a slot.
The same page can serve both, but the failure modes differ. Google rewards a page that ranks. An answer engine rewards a page it can quote without editing. In our window, Google AI Mode named the site 14 times and ChatGPT 9, while Claude, Gemini and Perplexity named it zero times, which tells you one page cannot be tuned to a single surface.
Long enough to answer the question and every obvious follow up, short enough that the answer is in the first paragraph. We do not publish a word count target, because we have not measured one that holds across surfaces.
A real FAQ does, when each question is one a guest actually asks and the answer stands on its own out of context. What does not work is a FAQ written to fill a schema block, because the answer only makes sense next to the question it sits under.
It does not force a citation. It makes your facts unambiguous, which is a different and cheaper job. The AI Readiness Check scores your site from 0 to 100 on access, structure and verifiability, so you can see which of the three is actually blocking you.
The pixel shows which of your pages receive the sessions that arrive from assistants, and prompt tracking stores the source list of every answer, so you can see which of your URLs the models picked. In the window above, the measured site supplied 73 of the 784 citations.
Not usefully. The ten booking and ticketing brands in that window took 144 citations between them, so copying their wording puts you in a queue behind pages the models already trust for exactly that phrasing.
They describe a measurement. The 7 domains behind them are our own travel websites rather than hotel clients, and it is one pass, so a single reading proves nothing. Use the mechanics here and get your own baseline before drawing a conclusion about your property.