Em dashes and AI writing: what 11,224 flagged sentences show

The em dash became the most argued-about punctuation mark on the internet somewhere around 2024, when people noticed chatbots scattering it through everything they wrote. Writers who had used it happily for decades started deleting it from their own drafts, worried about being mistaken for a machine.

SiteTell's ruleset has an opinion on this, and some numbers behind it. Across the 80 sites whose page text we rescored, 50 tripped the em dash rule at least once. That is 63%, which puts it level with the rule of three as the second most common structural pattern we measure. The more useful finding, though, sits in how the rule decides, and in where on a site the dashes pile up.

Does using em dashes mean AI wrote it?

No. A single dash, or a handful spread across a long page, tells you nothing about who wrote it. Novelists, journalists and copywriters have leaned on the mark for a very long time, and the rule was built with that in mind.

What it measures instead is rate. The check only runs on pages of 100 words or more, and it starts counting once dashes reach roughly one per 500 words. Below that, a page can use them freely and never see a flag. Above it, the penalty grows with density until it saturates at six per 500 words, the point where the mark stops punctuating prose and starts replacing it.

That design follows a note in the research we built the ruleset from: real writers use em dashes too, so the typical machine rate is the signal, never mere presence. It is also why the rule carries low severity, the lightest of the three levels the ruleset uses.

How does the rule count?

Each sentence containing a dash becomes its own flag, once a page is over the threshold. That was a change. Until 2026-09-12 the rule emitted one flag per page, and the result was a fix that looked broken: someone would correct one sentence out of eight, rescan, and find the flag still there, because the page-wide ratio had barely moved. Now every offending sentence is listed, so correcting all of them clears the rule, and correcting some visibly reduces the penalty.

The switch took the average from 24.2 flags per scan to 89.5. The penalty itself was held steady: the page-level severity is divided across the flagged sentences, so a page with eight of them costs the same as it did when it had one flag.

That explains the rule's instance count. It produced 11,224 flagged sentences on 2,214 pages, far more than any other structural rule, and the count mostly reflects how the rule reports rather than how bad the problem is. Share of sites is the steadier figure.

Where do the dashes cluster?

On homepages, by a wide margin. Per 1,000 words of rescored text, homepages carry 3.98 flagged sentences. Documentation follows at 2.28, and blog posts come in at 0.71, which makes the homepage rate about 5.6 times the blog rate. The homepage bucket spans 25 sites and no single site contributes more than 28% of its hits, so this is a spread pattern, not one site's quirk.

Our reading of why: homepages are where copy gets compressed. A hero line, a subhead, a row of feature tiles, each with a few words to land a point, and the dash is the fastest way to bolt a qualifier onto a fragment. Blog posts have room for full sentences, so the qualifier gets a clause instead.

Pages that trip this rule also lean slightly toward particular vocabulary. They are about twice as likely as an average page to carry seamlessly, a pattern spanning 13 sites, and 1.8 times as likely to carry landscape. Neither pairing is dramatic, which fits a rule whose trigger is formatting habit rather than word choice.

What does a flagged sentence look like?

Every pair below was written for this page, since quoting a real site could identify it. All of them are tested on each build of this site: the flagged side has to trip the em dash rule, and the rewrite has to pass every rule SiteTell can check on a single passage.

The dashes in that hero line are carrying an aside, a list of teams, that the sentence never needed. The rewrite drops the list and spends the space on a consequence a buyer would care about.

Pricing and about pages share the pattern: a dash introducing the punchline. A comma, or simply restructuring so the point comes first, usually reads better.

Paired dashes around an aside can nearly always come out. If the aside matters, it deserves its own clause; if not, it can go.

What about date ranges, versions and citations?

These are exempt from the rewrite, and the reason is a mistake we made. The first version of the rule flagged changelogs and pricing tables, where a dash between two dates or two version numbers is typography, not prose. The instruction our fix engine now follows says to leave that kind of dash alone and return the sentence unchanged.

The detector itself still counts every dash on the page, so a changelog dense with date ranges may cross the threshold. Each flag shows its sentence, which makes those easy to spot and ignore.

Is the em dash still an AI signal at all?

This is a fair question, and we asked it ourselves. In August 2026, reporting on a study of model output suggested that most major models had since been tuned to use fewer dashes than human writers do, with one exception. If true, a high dash rate would point at one model family at most, and a low rate would point at nothing.

We could not verify the underlying study directly, so we have not built anything on its specific claims. We did consider retiring the rule, and on 2026-09-12 decided to keep it. The deciding argument was about the copy that exists rather than the models being released: sites carry pages written over years, and a great deal of what is live now was generated when heavy dash use was the norm. Kept at low severity, the rule points at that copy without dominating anyone's score.

Whatever the models do next, the editing advice holds on its own terms. A page where every third sentence breaks mid-thought is harder to read, whoever wrote it.

How to fix a page that trips it

Go through the flagged sentences and ask what each dash is doing. Most fall into one of four jobs:

1. Introducing a punchline. Use a colon, or rewrite so the point comes first.
2. Wrapping an aside. Use commas or parentheses, or cut the aside.
3. Joining two complete thoughts. Use a full stop or a semicolon.
4. Separating two numbers or dates. Leave it; that is typography.

You rarely need to remove every dash. Getting a page back under one per 500 words clears the rule, and keeping the few that genuinely earn their place is good writing.