“Write long enough and AI will treat you as authoritative.”

That belief came over intact from the SEO era, and it now props up an entire market of content services promising three thousand words a piece. It’s wrong, and there’s a dataset large enough to say so directly.

The study measured position, not whether you get cited

Start with the definition, because this number has already been mangled several times over.

Ahrefs analysed 560,000 AI Overviews, pulled 1.67 million cited URLs out of them, and filtered down to 174,000 pages whose content could be extracted cleanly. What they measured was the relationship between word count and where a page sits in the AI Overview’s citation list. The Spearman correlation is 0.04.

0.04 is statistically nothing. In plain terms: when a one-thousand-word page and a five-thousand-word page answer the same query, length barely influences which one gets listed first.

Note what that does and doesn’t cover. It measured citation position, not whether a page gets cited at all. Plenty of write-ups render this as “word count has nothing to do with being cited,” which goes a step further than the data. What the data supports is narrower: among pages that already made the shortlist, length won’t move you up it.

That limit doesn’t weaken the practical conclusion, because the next number is more direct.

More than half of cited pages are under a thousand words

From the same dataset: 53.4% of pages cited in AI Overviews come in under a thousand words, and only 16% run past two thousand.

The length distribution of AI citation sources runs opposite to the belief. Short pages aren’t being filtered out — they’re the majority.

Which makes sense once you picture the mechanism. An AI doesn’t read your page end to end and then decide whether to recommend you. It lifts one passage and drops it into an answer. If a passage can be lifted, you have a shot; if it can’t, a hundred thousand words won’t help. Length has no role in that step. Where it does have a role is the step before — the page has to fit into the model’s context, and past a point the length works against you.

What does correlate: the text between two headings

The same research measured something else: how many words sit between one heading and the next.

Pages with 120 to 180 words between headings earned roughly 70% more ChatGPT citations than pages with fewer than 50.

This is a far more useful number than 0.04, because you can act on it directly. What it describes is a block big enough to make a complete point on its own, but not so big that a reader — or a model — loses the thread.

Too short and the information is incomplete: the AI lifts the block, finds a bare claim with no conditions, no figures and nothing to stand on, and goes to lift someone else’s instead. Too long and it’s diluted: eight hundred words under one heading, and the model can’t tell which part to take, so it skips the block entirely.

How this fits with “40 to 80 words per paragraph”

Our earlier piece on designing citable passages recommends keeping a single paragraph to 40–80 words. That sits at a different level from the 120–180 figure above, and the two shouldn’t be mixed up:

  • 40–80 words describes one paragraph — the unit an AI actually lifts
  • 120–180 words describes the space between two headings — the block that unit lives in

They nest cleanly: two or three 40–80 word paragraphs under one heading lands you in the 120–180 range. That isn’t a coincidence, it’s the same thing measured at two scales. A liftable paragraph needs to sit inside a block the right size to be found.

In practice it looks like this. The heading carries information (not “The deeper reason”). The first paragraph answers what the heading asks. The second adds conditions, figures or exceptions. The third closes on something that reads correctly on its own. Then a new heading.

So what should you cut

If you’re holding a stack of content written to hit a word count, cut these first:

  • Recaps. “As we mentioned earlier” at the top of section two is noise to a reader arriving from search and to an AI lifting a passage from the middle.
  • Restatement. Saying one thing three ways was keyword density in the SEO era. Now it’s dilution.
  • Section summaries. The symmetrical intro-body-summary structure with a wrap-up under every heading is itself a template signal, and the wrap-up is usually the previous paragraph reworded.
  • Abstract triads. “Faster, sharper, more complete” takes up word count and contains no liftable fact.

If the piece loses half its length, that isn’t a loss. Concise doesn’t mean shallow — the density goes up, and every passage that comes out carries facts and conditions with it. That’s what citable looks like.

The hard part isn’t cutting, it’s knowing whether the cut worked

Everything above you can start on today; an hour per article is realistic.

The awkward part comes after. How do you know it worked? Rankings won’t move, traffic may not shift for a while, and the only thing you can actually observe is whether AI engines start lifting your passage when they answer related questions. That means asking, round after round, recording which questions cited you and which passage they took — and because AI answers carry real randomness, a single result proves nothing.

For one article, that loop runs for weeks before it tells you anything. Across a few dozen pages, it’s a different order of work altogether.

If you want to know whether your content can be lifted as it stands, we can take a look first.

Further reading