Writing for LLMs Without Writing Badly for People

The one rule underneath all of it

Write so that any single paragraph still makes sense with everything around it removed. That is the entire discipline. Every specific piece of advice below is a consequence of it.

The reason is mechanical. A system composing a short answer does not read your page and summarize it. It takes passages. A paragraph that depends on the three above it to be intelligible arrives at the destination as nonsense, and will not be used.

This sounds like it should produce stilted writing and it does not, which is the genuinely useful thing about it. Self-contained paragraphs are what good reference writing has always done. The habit being corrected is a marketing habit rather than a writing one.

What does produce bad writing is the other interpretation of this advice, which is writing for the machine as though it were a separate audience with a separate dialect. That produces pages of question headings and one-line answers that read as thin to people and gain nothing.

Lead with the answer, then earn it

The most valuable change available on most sites is moving the conclusion to the top, and it costs nothing because the conclusion already exists further down.

Commercial pages are built to hold attention. Context, complication, then resolution. That structure serves a reader who arrived and is going in order, and it hides the sentence worth quoting behind three paragraphs of setup.

The alternative is to state the claim plainly in the first sentence or two and spend the rest of the page earning it. This is how good reference writing works, and readers do not object, because a person who wanted the answer has it and a person who wanted the reasoning reads on.

The objection is usually that the build is doing persuasive work. Sometimes it genuinely is, and the honest resolution is to keep the build and add a plain statement above it rather than to choose. A short, flat answer followed by the argument serves everybody, including the extractor.

Make every sentence carry its own subject

This is the most mechanical habit on the list and the fastest to learn.

Sentences beginning with "this", "it", "they" or "that" point backwards. In place, the reader knows what is meant. Lifted, the reference is gone. Naming the subject again costs three words and makes the sentence portable.

Comparisons need both halves. "Considerably faster" invites the question faster than what, and if the other half was named a paragraph ago, the extracted version answers nothing.

Figures need their unit and their basis. A number with no indication of what it counts is not a fact, it is a decoration, and it is worse than useless when it turns up in an answer detached from the sentence that explained it.

None of this makes prose clumsy if it is done at the right level. You are not repeating the subject in every clause, you are making sure each paragraph opens by establishing what it is about.

One idea per paragraph, named in the heading

Structure does as much work here as sentence construction, and it is easier to fix.

A paragraph holding two ideas is a paragraph that can only be used whole, and whole is often too long. Splitting them gives two usable units where there was one awkward one, and it reads better besides.

Headings should state what the section establishes rather than tease it. A heading written as a hook makes a reader work out what is below it, and gives anything else no information at all. A heading that names the conclusion helps a skimmer and helps an extractor find the relevant part.

Use a table where a table is genuinely clearer. Comparisons across several dimensions are painful in prose and obvious in a grid, and a well-formed table is unusually easy to reuse accurately. Do not use one where a sentence would do, because a table of two rows is a sentence somebody has made harder to read.

Say something only you can say

This is the part that cannot be systematised, and it is the part that decides whether a passage is worth reaching for.

A sentence that could appear on any competitor's site adds nothing to an answer. There is no reason to prefer your phrasing of a generic claim over anybody else's, so the choice falls to whoever is better established.

What makes a passage worth quoting is a real constraint you have hit, a distinction you draw that others do not, a trade-off you are willing to name, or a number you can stand behind. All of those come from having done the work, which is why this cannot be outsourced to somebody who has not.

Conceding something is the most reliable version of this. A page with no trade-off in it reads as promotion, and promotional passages are poor material for an answer trying to be useful. Naming the case against your own position is what makes a source sound like a source.

The advice that produces bad pages

Several tactics circulate under this heading and each one makes pages worse without helping.

Stuffing question headings. A page of thirty question headings with one-line answers reads as a quiz and is thin by any measure. Questions as headings are fine where the section genuinely answers a question and terrible as a template.

Repeating the key phrase. Density does nothing except make the writing worse, and worse writing is less quotable, so this is actively counterproductive.

Writing to a word count. Padding adds sentences that dilute the ones worth lifting. If the subject is finished at 900 words, adding 1,100 more makes the page harder to extract from, not easier to retrieve.

Explaining what the page will cover. An opening paragraph describing the structure of the page is a paragraph containing no claim. It is the least useful thing that can occupy the most valuable position.

What to change in house style

Most of this can be encoded as house rules rather than relearned by each writer, which is how it survives staff changes.

Require the first two sentences to answer the title. This is checkable by anybody in ten seconds and it is the highest-return rule available.

Ban paragraph-opening pronouns. It sounds pedantic and it eliminates most extraction failures on its own.

Require every comparison to name both sides and every figure to carry its unit. Also checkable, also mechanical.

Require one concession per piece. This is the rule that most improves the writing and the one most likely to be argued about, because it asks a marketing page to say something that is not flattering.

What not to encode is length. A floor makes sense as a signal that a subject has been treated properly and a target produces padding, which is the thing you are trying to avoid.

Does any of this hurt the page for people

This is the fair objection and it deserves a direct answer rather than a reassurance.

Most of it helps. Leading with the answer, one idea per paragraph, honest headings and self-contained sentences are all standard advice for clear writing and were before any of this existed. A page rewritten this way is easier to read, not harder.

One thing genuinely costs you something. Stating the conclusion up front removes the build, and for some subjects the build is doing real persuasive work. The resolution is to state it plainly and then build anyway, which is more words than either approach alone.

The other cost is that specificity invites disagreement. A page that names a trade-off can be argued with in a way a page of general claims cannot. That is a real exposure and it is also the thing that makes the page worth citing, so it has to be accepted rather than managed away.

The honest trade-off

Writing this way takes longer and it takes a different kind of effort. Producing a self-contained, specific, conceding paragraph is harder than producing a fluent one that circles, because you have to know what you actually think.

It also produces fewer pages for the same budget, and fewer pages is the correct direction here, which is difficult to explain to anybody measuring output. Depth on one subject beats coverage of many, and depth means writing less, better, about a narrower thing.

The compensation is that the downside case is mild. If none of this moved your visibility at all, you would still be left with a clearer site, which is not true of most tactics in this area.

There is one more cost worth naming, which is organizational rather than editorial. These rules make writing checkable, and checkable writing invites review. A house style requiring the first two sentences to answer the title gives everybody in the approval chain a legitimate reason to send a draft back, and that slows publishing. It is a real cost and it is worth paying, because the alternative is a rule nobody applies and a set of pages that quietly drift back to circling before they commit.

Editing what you already have

Most of the value here is in existing pages rather than new ones, and the edit is smaller than people expect.

Start with the ten pages that carry commercial weight. Not the ten most trafficked, the ten that matter to the business, because those are the ones where being quoted is worth something.

For each, read only the first two sentences. If they do not answer the question in the title, find the sentence further down that does and move it up. You are relocating your own conclusion, not writing anything new, and this alone is usually an afternoon for the whole set.

Then scan for paragraph openings that begin with a pronoun and give each one its subject back. This is mechanical enough to hand to somebody without deep subject knowledge, and it eliminates most extraction failures.

Then look for a claim on each page that only you could make, and if there is not one, that is the finding. A page with nothing specific in it will not be reached for however well it is structured, and the fix is knowledge rather than editing. Our page on how LLMs choose sources covers why specificity carries so much weight at the selection gate.

How to tell whether the writing worked

Editorial work is unusually measurable here compared with most writing, and that opportunity is regularly wasted.

Assistants that show their sources will tell you which passage was taken. If you rewrote an opening and that opening subsequently appears as the quoted material, you have direct evidence that a specific edit landed. Very little else in content work gives feedback that clean.

So the sequence is worth respecting. Record what is currently being quoted from your pages, if anything. Make the edit. Watch whether the quoted passage changes to the one you wrote.

What you cannot do is read this from your analytics, because the visit that follows a citation is indistinguishable from any other direct visit. Anybody promising attribution reporting for this is describing something the plumbing does not currently support. Our piece on citation tracking covers what can honestly be observed.

Give it weeks rather than days, and repeat the reading. A passage that was quoted once may not be the next time, and the thing you are watching is frequency across many readings rather than a single event.

Bottom line

One rule underneath everything: any paragraph should make sense with the rest removed. Lead with the answer, keep one idea per paragraph, name the subject in every sentence, and put both halves of every comparison in.

Then say something only you can say, and concede something real. Specificity and honesty are what make a passage worth reaching for, and no structural technique substitutes for either.

Do not write for the machine as a separate audience. There is no dialect to learn. There is clear writing, which happens to survive extraction, and unclear writing, which happens not to.

If you take one thing from this page, take the extraction test. Copy a paragraph out of your own site into an empty document and read it as a stranger would. Almost everything worth fixing shows up in that thirty seconds, and it needs no tooling, no budget and nobody's permission.