Pew Research Center analyzed nearly half a million English language web pages pulled from the Common Crawl archive, run through the Open Pangram AI detector. The headline number is that roughly 10 percent of pages sampled in July 2026 show signs of AI authorship. Restrict that to pages published since ChatGPT launched in late 2022, and the share climbs past a third.
That is a real, measured jump, not a guess about how the internet feels lately.
Where the AI text actually concentrates
The study broke results down by domain type, and the split is not close.
| Domain type | Share showing AI text |
|---|---|
| .com | ~10% |
| .org | 4.6% |
| .edu | ~1% |
| .gov | ~1% |
Commercial sites run roughly ten times higher than educational or government ones. That tracks with incentive rather than access. Anyone can run a model against a keyword list and publish, and a .com site has a direct reason to do exactly that. A university department or a government agency mostly does not.
The tells the study measured are the same ones an editor would catch
Pew also tracked how web writing style shifted since 2023, independent of whether a page was flagged as AI written.
- Em dash use roughly doubled
- Oxford comma use is up 63 percent
- Words models overuse, things like delve, pivotal, and landscape, more than doubled in frequency
- Constructions like "it's not just X, it's Y" nearly tripled
None of that is a smoking gun on its own. A human editor reaches for an em dash too. But stacked together, across half a million pages, the shift is large enough to show up in aggregate style statistics, which is a different kind of evidence than one suspicious paragraph.
What the study cannot tell you
The researchers are upfront about a real limit. Their detector cannot reliably separate fully automated text from writing that started human and got AI-assisted editing, a rewrite pass, or machine translation. A page flagged as showing AI signs might be entirely generated, or it might be a person's own draft run through a tool for polish. The 10 percent and 33 percent figures describe pages with AI fingerprints, not pages proven to have no human author at all.
That distinction matters more than it looks. It means the study measures the spread of a writing style, not the disappearance of writers.
What this means for a business publishing content
Three things follow directly from where the numbers concentrate.
Being on a .com domain does not exempt a page from scrutiny, it makes scrutiny more likely. Readers and search engines are calibrating against a domain type where AI text already runs at ten times the base rate, so a thin or generic page has less benefit of the doubt than it did in 2022.
The same stylistic tells the study measured are worth checking in your own drafts before publishing, not as a purity test but because they correlate with the generic, low-specificity writing that both readers and Google's helpful content systems tend to penalize. An em dash is not the problem. A page that could have been written about any company in any industry is.
Primary detail is the actual differentiator. The study cannot tell fully automated text apart from AI-assisted text, but a reader usually can, once a page includes something a generic pass could not produce: a specific number from your own data, a screenshot from your own test, a name of someone who actually said the quoted thing.
The rule worth keeping
AI-written text is not rare anymore, it is roughly one in three pages published since the tools became mainstream, concentrated hardest exactly where content gets published fastest and cheapest. The pages that hold up are not the ones without an em dash. They are the ones with something in them a template could not have produced on its own.