The short version
- Mostly no difference. Clear, well-structured, factually specific writing serves both — which is convenient and slightly anticlimactic.
- The genuine differences are structural: self-contained answers, explicit statements over implication, and attributed facts.
- What does not help: writing "for the algorithm", keyword stuffing, or any of the proprietary "LLM readability" metrics being sold.
- The strongest differentiator is content a model cannot synthesise — your data, your tests, your experience.
A genre of advice has grown up around writing specifically for language models, and much of it recycles SEO folklore with new terminology. The honest position is that good writing for a knowledgeable reader is close to optimal for a machine trying to extract facts from your page, with a handful of specific exceptions worth knowing.
What is genuinely the same
- Clarity. Ambiguous prose is hard for everyone. A sentence that could mean two things is a liability regardless of the reader.
- Structure. Headings that describe content help a person skim and a machine locate.
- Accuracy. Wrong information is wrong for both audiences, and being wrong in an extractable way is worse than being wrong vaguely.
- Depth. Superficial coverage satisfies neither.
- Specificity. "$129/mo" is more precise than "affordable" for a reader and for an extractor.
What genuinely differs
The last row is the most practical and least discussed. If a paragraph begins "It also includes unlimited seats", a reader knows what "it" means from context. Lifted out of the page, that sentence is useless. Naming the subject — "Clearscope also includes unlimited seats" — costs two words and survives extraction.
What does not help
- Keyword stuffing. No more effective here than it has been in search for a decade.
- "LLM readability scores". Vendor-invented metrics with no published relationship to how any assistant selects sources.
- Writing in an artificially simple register. These systems handle complex prose fine; dumbing down loses precision.
- Question-shaped thin content. Twenty FAQ entries with no substance fails both audiences.
- `llms.txt` as an optimisation lever. See the comparison — it is a proposal, not a standard.
The structural habits worth adopting
Three changes cover almost all of the genuine delta, and none of them costs anything in readability. Name the subject in each claim rather than relying on a pronoun that has travelled two paragraphs. Put the direct answer before the elaboration. Attach a date and a source to any figure, which Google’s guidance on people-first content has effectively encouraged for years on entirely separate grounds.
The fourth habit is structural rather than stylistic: use headings that describe content rather than tease it. "Pricing compared" is locatable; "The elephant in the room" is not. Combined with valid structured data, that gives any system parsing your page an unambiguous map of what is where — which is the same thing it gives a reader skimming on a phone.
A useful way to check whether you have done this well is to take any paragraph from your page, paste it somewhere with no surrounding context, and read it as a stranger. If it still makes sense and still identifies what it is talking about, it will survive extraction. If it depends on the three paragraphs above it to be intelligible, it will be quoted badly or not at all — and that is a fixable writing problem rather than an algorithmic one.
The habit that serves both, demonstrated here
Every factual claim on this blog carries a source and a date. That exists because readers should be able to check figures — and it happens to be exactly what makes a claim safe for a machine to repeat. When advice for the two audiences converges this neatly, it is usually because the underlying property is just honesty.