When ChatGPT or Perplexity answers a question using your docs, it doesn't reproduce the page. It lifts a sentence or two and builds its answer around them. So the unit that matters for showing up in AI answers isn't the page. It's the sentence.
That changes how you write the important parts.
Put the answer in one self-contained sentence
A model extracts cleanly when the answer doesn't depend on the sentence before it. If your direct answer is "It expires after that period," the model has to also pull the previous sentence to know what "it" and "that period" are. Often it won't, and you get misquoted or skipped.
Write the load-bearing sentence so it stands alone:
That sentence carries its own subject ("ReadMe API keys"), its own claim ("do not expire automatically"), and the qualifier a model needs to not overstate it. Lift it out of the page and it's still true and still complete. That's what makes it quotable.
Lead with the answer, then explain
Models weight the first sentence under a heading heavily, because that's where direct answers usually live. If your section opens with three sentences of context before the actual answer, the model may quote the context and miss the point.
Front-load it. State the answer in the first sentence, then spend the rest of the paragraph on the nuance, the caveats, and the example. Humans skim this way too, so you're not trading reader experience for machine readability.
Survive the chunk boundary
Retrieval systems split your page into chunks before a model ever sees it, and they don't split where you'd want. If your answer and its qualifier sit on opposite sides of a chunk boundary, the model gets half of it. Keeping each answer self-contained is the same discipline that makes a page survive chunking: no sentence should need a different chunk to make sense.
How ReadMe puts this together
Getting quoted is mostly writing, but it's also structure. The same work that makes a page cite-able by an AI engine makes individual sentences extractable: clear headings, complete examples, and an llms.txt that tells models where to look. ReadMe publishes that scaffolding for you so the writing is the part you focus on.
Write the sentence you'd want quoted, then check whether it survives being pulled out of the page. If it does, it'll survive a model too. Start a project to see how your docs read to one.
