Pew Finds AI Signals on 35% of Dated Post-ChatGPT Webpages

Commercial domains led the rise

The increase was not evenly distributed across domain types. In 2026, roughly 10% of .com pages showed significant AI-authorship or editing signals, compared with 4.6% of .org pages. The corresponding rates for .edu and .gov pages were each about 1%.

Before ChatGPT’s release, the four domain groups had broadly similar rates. The later divergence is consistent with faster adoption of AI-assisted production in commercial web publishing, although the study’s detector results do not establish why the rates differed.

The web’s writing style changed, too

Pew also tracked stylistic features that became more common after ChatGPT’s release. These included:

• Em dashes
• Oxford commas
• Words such as “delve,” “interplay,” “tapestry” and “testament”
• “Negative parallelism,” a construction such as “it’s not just X, it’s Y”

The measured frequency of these markers rose across the dataset over time. But they are statistical patterns, not fingerprints. A human writer can use an em dash or one of these words without using AI, and a page can contain AI-assisted text without displaying every listed feature.

Why the 35% figure needs context

The most important caveat is that detection is probabilistic. Pew treated a detector score of 0.2 or higher as a meaningful sign of AI use, while acknowledging that detectors can misclassify individual documents. The method is therefore more useful for identifying broad changes across a large, randomly sampled dataset than for declaring that a particular article was written by AI.

It also measures a continuum of human and machine involvement. “AI-authored” in this context can include substantial editing or assistance, not just fully automated generation. That distinction matters for publishers, researchers and readers trying to interpret the findings.

A web increasingly shaped by automated systems

Pew’s results describe the production and style of web content. Separate reporting based on Cloudflare data found that automated systems accounted for about 57% of activity on its network in June 2026, compared with about 43% for humans. That figure concerns measured traffic on Cloudflare’s network—not every kind of internet activity—and it does not show that the pages flagged by Pew were authored by those bots.

Taken together, the findings point to a broader change: online information is increasingly produced with AI, discovered or processed by automated systems, and consumed in environments where humans are no longer the only meaningful audience. The evidence supports a substantial shift in how the web is made and navigated, while leaving the authorship of any single page an open question.