How Much of the Internet Is AI Generated? Pew's Number
How much of the internet is AI generated? As of July 2026, about 10% of all sampled web pages show significant signs of AI authorship, and among pages published after ChatGPT launched in November 2022, that figure rises to over a third.
How much of the internet is AI generated? As of July 2026, about 10% of all sampled web pages show significant signs of AI authorship, and among pages published after ChatGPT launched in November 2022, that figure rises to over a third. Those numbers come from a Pew Research Center analysis published on 20 August 2026, and they are the closest thing to a measurement anyone has produced at this scale.
They are also easy to misread, which matters more than the headline does.
How much of the internet is AI generated, by the numbers
Pew pulled roughly 490,000 English-language pages from the Common Crawl archive spanning January 2021 to July 2026, then took a random sample of 10,000 pages from the July 2026 snapshot and ran them through Open Pangram, a machine-learning detector. The full write-up is on the Pew Research Center site.
Slice | Share showing AI authorship |
|---|---|
All sampled pages, July 2026 | about 10% |
Pages published after November 2022 | over one third |
.com domains | about 10% |
.org domains | 4.6% |
.edu domains | about 1% |
.gov domains | about 1% |
Pew is candid about the limitation. Detection models, in its words, sometimes misclassify documents written by humans as showing signs of AI authorship, and the reverse. That caveat is not boilerplate. It is the whole methodology.
Why a detector number is not a census
A detector does not read a page and know who wrote it. It scores the text for statistical patterns that appear more often in machine-generated writing than in human writing: certain function-word frequencies, low burstiness in sentence length, particular collocations. A human writer who edits tightly and writes in short declarative sentences can trip it. An AI draft that a person rewrote heavily can slip past it.
At the level of a single page that makes the score close to useless. At the level of 10,000 randomly sampled pages, the errors in both directions partly cancel and the aggregate becomes directionally informative. That is the correct way to hold this number: a good estimate of a trend, a bad basis for accusing any individual page.
It is the same failure mode covered in how to tell if an AI answer is hallucinated, where a confident-sounding output gets treated as a verdict rather than a probability.
The domain gap is the interesting part
The headline everyone quoted was the one-third figure. The more useful finding is the spread between top-level domains: commercial sites sit around 10%, while .edu and .gov sit near 1%, roughly a tenfold difference, with .org in between at 4.6%.
That gap tells you where AI writing concentrates, and it is exactly where you would expect: pages produced under commercial pressure, at volume, where the marginal cost of another page is the thing being optimised. Institutional and government pages face neither the same volume incentive nor the same publishing cadence.
For anyone publishing commercially, that is the honest read. You are operating in the part of the web where the density is highest, which means the average page you compete with is more likely to be machine-drafted than the average page anywhere else.
What this changes for people publishing with AI
Not much about tactics, and quite a lot about expectations.
Detector scores are not a compliance signal. Nobody has a rule that says a page scoring above some threshold gets demoted. Google's stated position remains about quality and intent, not authorship method, which is covered in more detail in does Google penalize AI generated content.
Disclosure obligations are a separate question entirely. Whether you have to say a page was AI-assisted depends on jurisdiction and context, not on a detector, and that is unpacked in do you need to label AI generated content.
Provenance signals are moving faster than detection. Watermarking and content credentials attach at generation time rather than being inferred afterwards, and their limits are set out in can AI text watermarks be removed.
If a third of new commercial pages look machine-drafted, differentiation is the only defensible position. Not human authorship as a badge, but something on the page that a generic draft could not contain.
Full disclosure, since it would be strange not to say it here: this blog is written with AI assistance. That is precisely why the differentiator gate matters, and why a number like Pew's is worth reading carefully rather than quoting.
Frequently asked questions
Is a third of the internet AI generated?
No. About 10% of all sampled pages showed AI authorship signals. The one-third figure applies only to pages published after November 2022, which are a subset of the web, and it is a detector estimate rather than a count.
How accurate are AI content detectors?
Accurate enough to be useful across thousands of documents, unreliable on any single one. Pew states plainly that its detector misclassifies in both directions. Treat an individual score as weak evidence, never as proof.
Does Google use AI detectors to rank pages?
Google has never said it does, and its published guidance targets unhelpful, scaled content rather than authorship method. The August 2026 spam update continued that framing.
Where did Pew get the pages it analysed?
From Common Crawl, an open archive of crawled web pages. The sample was English-language only, which means the figures describe the English web rather than the whole of it. TechCrunch's coverage has the summary if you want the shorter version.
Keeping track of studies like this without drowning in launch posts is its own skill, and there is a method for it in how to keep up with AI news.
How did this land?
About the author
Senior Editor, AI & Product
Cecilia leads the Swarmz editorial desk. She has spent a decade turning complex AI and product topics into writing people actually finish, and she owns the blog's quality bar.
