Pew Finds AI Signals on 35% of Web Pages Published Since ChatGPT

Pew Research Center’s August 2026 analysis offers one of the clearest recent estimates of how extensively AI has entered online publishing: meaningful signs of AI authorship or substantial editing appeared on 35% of sampled English-language webpages published after ChatGPT’s public release. In a broader July 2026 snapshot that included older pages, the share was 10%. 15

The distinction matters. Pew’s system identified patterns associated with AI writing or editing; it did not establish that AI generated every flagged page, or determine how much AI assistance a page received. 1

What Pew studied

Pew drew 10,000 English-language webpages from each of 49 Common Crawl snapshots collected between January 2021 and July 2026, producing a total sample of 490,000 pages. Common Crawl is an archive of publicly accessible web content, so the sample does not represent every page on the internet equally; material behind paywalls or login screens is likely less visible in the archive. 3

The researchers extracted body text from the pages and analyzed it with Open Pangram, an AI-detection system designed to identify signals associated with AI authorship or substantial editing. The analysis treated the result as a continuum rather than a simple human-versus-machine classification: a page could show signs of mixed human and AI production as well as more extensive machine generation. 13

The headline results: 10% overall, 35% after ChatGPT

A random sample of 10,000 pages collected in July 2026 found significant AI-authorship signals on about one in 10 pages. That sample included pages published before ChatGPT, which lowered the overall rate. 15

Pew then filtered the sample to pages with publication dates after ChatGPT’s Nov. 30, 2022 public release. Within that dated subset, the share rose to 35%, or more than one-third. 1

That comparison should not be read as a precise census of every page published since ChatGPT launched. Only a minority of webpages contained a usable publication-date field, so the 35% estimate applies to pages Pew could date and compare—not necessarily to all post-ChatGPT pages. 3

Commercial domains showed more AI signals

The prevalence varied substantially by top-level domain. In Pew’s 2026 measurements:

• .com: roughly 9%–10% of pages showed significant signals.
• .org: about 4.6%.
• .edu: about 1%.
• .gov: about 1%.

The pattern puts .com pages at approximately twice the .org rate and about 10 times the .edu and .gov rates. 4

These figures describe differences in the sampled pages, not a finding that every commercial publisher uses AI or that educational and government websites are unaffected. They do show that the strongest measured concentration was on the commercial web. 4

How this relates to Cloudflare’s bot-traffic finding

Cloudflare-related reporting in 2026 said automated bots and AI agents accounted for roughly 57%–57.5% of measured requests to HTML content, compared with about 42.5% for humans. 1718

That statistic and Pew’s estimate describe different layers of the web:

• Pew measured page content: whether text contained signals associated with AI authorship or substantial editing.
• Cloudflare measured requests: whether traffic reaching HTML pages appeared to come from automated systems or people.

The two findings are therefore complementary, not interchangeable. A bot can crawl a page written entirely by a human, and a page showing AI-authorship signals can still be read mostly by people. Neither study proves the other statistic. Taken together, however, they point to a web increasingly shaped by both machine-produced content and machine-driven access. 11718

The study’s main limitations

Pew cautioned that AI detectors can misclassify human-written text. A positive result is consequently best understood as a signal, not definitive attribution of authorship. The method also cannot reliably distinguish between wholly AI-generated writing and lighter forms of AI assistance, such as rewriting or editing. 1

The sample has additional boundaries:

• It covers English-language webpages rather than the global web.
• It relies on publicly crawlable pages in Common Crawl.
• The post-ChatGPT comparison uses the subset with embedded publication dates, which is not necessarily representative of all newer pages. 3

Those limitations make the 35% figure an estimate of detectable AI-associated signals in a specific sample—not a final answer to what percentage of the entire internet was written by AI.

Linguistic patterns also changed after ChatGPT

Pew examined several additional indicators as a separate way to track changes in web language. Among pages published after ChatGPT’s release, the researchers observed increases in the use of em dashes, Oxford commas, negative parallelism, and a group of words associated with AI-generated prose, including “delve,” “crucial,” “landscape,” “pivotal” and “underscore.” 2

These patterns are not authorship proof either. A human writer can use the same punctuation or vocabulary, and writing conventions can spread for reasons unrelated to AI. Their value is as corroborating evidence of a broader shift in online language, alongside the detector-based results. 2

The practical takeaway

Pew’s strongest finding is not that 35% of the internet is definitively AI-written. It is that, among the dateable English-language pages in its sample, more than one-third of post-ChatGPT pages showed meaningful signals consistent with AI authorship or substantial editing. The broader July 2026 snapshot found those signals on 10% of all sampled pages, while commercial domains showed considerably higher rates than .edu, .gov or .org pages. 134

The results are best used as a measure of a changing publishing environment—and a reminder that content provenance cannot be established from an AI detector score alone.