Five things marketers get wrong when writing for both humans and AI

Why the writing itself is now the variable

When the Princeton team that coined Generative Engine Optimization tested what actually moved a page into AI answers, the levers that worked were ordinary writing choices, not markup.

Across nine tactics and roughly 10,000 queries, adding relevant statistics lifted citations by about 41%, adding quotations by about 28%, and improving plain fluency helped too. Keyword stuffing — the reflex inherited from a decade of search — did nothing, and sometimes hurt.

That points somewhere unfashionable: the pages that get quoted tend to be the ones that were simply written better. Clearer claims, real numbers, plain sentences. So most of the gap is not a tooling problem. It is a writing problem, and it tends to show up in five recurring places.

1. Saving the answer for the end

Marketers are trained to build toward a reveal. A reader who skims, and a model that extracts, both tend to leave before the reveal arrives.

An answer engine lifts the passage that most directly resolves the question; it does not wait for paragraph nine to see where the argument lands. The fix is the oldest one in journalism: state the answer in a clean, self-contained opening sentence, then earn attention with the detail beneath it. This is not dumbing down. It is putting the conclusion where both readers look first — and it happens to be the same discipline that makes a page easy to quote.

A useful test: read only the first sentence of each section. If those sentences, on their own, already answer the question a reader arrived with, the page is built the right way round.

2. Writing claims too vague to lift

“We deliver best-in-class results” is unquotable. There is nothing in it a model can extract and stand behind.

A claim becomes liftable when it is specific, self-contained, and checkable — when the sentence still makes sense pulled out of the page and dropped into an answer. The Princeton finding that statistics raised citations by about 41% is, read closely, a finding about specificity: a number is a claim a model can carry without losing meaning. “Faster onboarding” carries nothing; “onboarding in two days instead of two weeks” carries the whole point.

This rarely means writing more. It usually means writing one sentence that could survive being quoted, then letting the rest of the paragraph support it. The vague version asks the reader to take the writer’s word; the specific version hands over something the reader, or the engine, can repeat.

3. Naming things loosely

A page that calls the same thing a “platform” in one line, a “tool” in the next, and a “solution” in the third is harder to summarize than it looks.

Models build a working picture of what a thing is from how consistently it is described. Loose naming, the kind a human reader glides past without noticing, leaves that picture blurry — and a blurry subject is one an engine tends to describe less confidently. Choose the noun for what you are and the category you sit in, then use them the same way on every page. The first time something appears, say plainly what it is, in the form “X is a [category] that [does what, for whom].”

This is the least glamorous item on the list and possibly the highest-leverage. It costs nothing, no human reader will ever complain, and it quietly removes the ambiguity a model would otherwise have to guess its way through.

4. Letting the number drift

A headline figure that reads “40%” on the website, “around half” in a deck, and “up to 45%” in an interview is, to a person, just rounding. To a model reading all three, it is noise.

When the same claim appears in three slightly different shapes, an engine has to decide which to trust, and a figure it has to adjudicate is one it tends not to repeat. Consistency here is unusual in that it sits entirely within a writer’s control. The honest move is to fix the canonical number and its wording once, then use that exact phrasing everywhere it appears.

5. Writing for the reader who has all day

The deepest habit is structural: building a page that only rewards someone who reads every word in order.

Almost no one reads that way now, and no engine does. A page works harder when its meaning survives skimming — when headings are real statements rather than labels, when the first line of each section can stand alone, when one idea sits in one paragraph instead of three braided together. None of this is a concession to machines. It is what a busy human wanted all along; the engine just makes the cost of ignoring it harder to hide.

The quiet test for the whole list: could a reader who skims, and a model that extracts, both leave with the right takeaway? When the answer is yes, you have resolved what is sometimes called the Dual Legibility Tension — the page a person reads and a machine can quote, written as one job rather than two.

What the evidence rewards, in one line

The research is consistent and faintly old-fashioned: it rewards substance a reader can verify, not tricks.

What measurably moved AI citations (Princeton GEO study, 9 tactics across ~10,000 queries)

None of the five fixes is exotic. Lead with the answer. Make the claim specific. Name things the same way every time. Lock the number. Write so the meaning survives a skim. What makes them hard is that they cut against habits built for a different reader — the slow build, the soft superlative, the page that only pays off in full. The writers pulling ahead in AI answers are not the ones with a hidden tactic. They are the ones who stopped writing for two audiences and started writing for the one reader who happens to be both.

Frequently asked questions