Writing content a language model will actually quote
Models lift sentences, not pages. That single fact changes how you structure an article, where you put the answer, and why buried conclusions never get cited.
Key takeaways
- Write the first forty to sixty words as a self-contained answer that survives being lifted alone.
- Phrase headings as questions and answer immediately beneath, since retrieval segments pages around headings.
- Replace adjectives with explicit figures, prices and dates, which had measurable citation effects where formatting alone did not.
- Name the brand inside the claim, because only what is inside the quoted sentence travels with it.
A language model composing an answer retrieves passages and lifts short spans from them. Content that gets quoted therefore shares a structure: the claim arrives early, it is self-contained, it carries a figure and a source, and the brand is inside the sentence rather than attached to it. Content that buries its conclusion is not quoted, however good the conclusion is.
The opening paragraph does most of the work
Write the first forty to sixty words as if they were the only part anyone reads, because for a large share of your audience that is now true. It should answer the question in the title directly, without preamble, and remain true if you delete everything around it. The common failure is an opening that sets up the topic instead of answering it. In a search-results world that cost you a little engagement; in an answer-engine world it costs you the citation entirely.
Structure for passage retrieval
- Headings phrased as the question, with the answer immediately beneath. Retrieval segments a page around headings, so a heading and its first sentence often travel together.
- One idea per paragraph. A paragraph making three points cannot be quoted for any of them without dragging in the other two.
- Lists for enumerable things and prose for arguments. A list of four items is easy to lift accurately; an argument compressed into bullet points loses the reasoning that made it credible.
- A short summary block of quotable facts. Explicit, standalone statements with numbers are the highest-value few sentences on the page.
Specifics beat adjectives
Comprehensive, cutting-edge and world-class are unquotable, because they assert nothing a model can stand behind. Replace every one with a fact. Not fast delivery but delivery in nine working days. Not affordable but from 1,800 dollars a month. Not extensive experience but forty projects in the Gulf since 2021. Explicit prices and dates were among the factors with measurable effects in the 2026 citation study, while formatting changes alone had weak ones.
The same principle governs statistics. A number without a source, a date and a unit is unusable to a system that has to attribute what it says. Give all three, inline, in the sentence containing the claim.
Say your own name
When a model reproduces a sentence, everything inside that sentence travels with it and everything outside does not. A claim written as our approach reduces response time by half gives the reader nothing to attribute. The same claim naming the company does. This feels awkward to write more than once or twice per article, and once or twice is the right frequency.
What not to do
There is a category of technique aimed at manipulating retrieval rather than earning it: hidden text, instructions addressed to models embedded in pages, keyword-stuffed passages engineered for extraction. Benchmarks now exist specifically to detect this behaviour, providers treat it as spam, and the reputational downside if you are caught is worse than the visibility upside if you are not. Write for the reader and structure for the machine; do not write for the machine.
Replace every adjective with a fact. Adjectives cannot be quoted, because they assert nothing.
Sources