Pew finds 35% of post-ChatGPT web pages were written by AI

Not long ago, Cloudflare dropped the news that bots officially generate more web traffic than actual humans. If that felt a bit dystopian, Pew Research just provided the punchline: a huge chunk of what those bots are reading was probably written by other bots in the first place.

According to a new study from Pew, more than a third of all web pages published since ChatGPT launched in November 2022 show clear signs of being written or heavily edited by AI.

The .com domain became an AI text factory

To figure out just how synthetic the open web has become, Pew analyzed nearly half a million English-language pages from the Common Crawl archive spanning the last five years. They ran the text through detection models from Open Pangram.

At first glance, a random snapshot of 10,000 pages from July 2026 showed that only about 10% had significant AI authorship marks. But that broad net caught millions of legacy pages that existed long before generative AI was even a thing.

Once researchers stripped away the pre-ChatGPT archive and isolated pages created after late 2022, the number skyrocketed to 35%.

Where that content lives tells an even bigger story. Commercial .com domains showed AI authorship rates roughly 10 times higher than .edu or .gov sites, which hovered around a clean 1%. Non-profit .org domains sat quietly at just 4.6%.

Pew also flagged stylistic giveaways that exploded alongside these tools. Tell-tale syntactical habits like heavy em-dash usage, the Oxford comma, and contrast clichés like "it is not X, it is Y" have visibly flooded modern publishing.

While automated detectors like Pangram aren't immune to false positives, the scale of this dataset makes the overarching trend undeniable: generative AI has fundamentally reshaped web publishing in less than four years.

Why it matters

For regular internet users, this marks the death of casual browsing trust. When more than one in three newly minted articles is machine-spun, that product review or recipe guide you just clicked on is increasingly likely to be synthetic filler built purely to farm search impressions. Finding genuine human perspective now requires deliberate filtering.

For the AI industry itself, this is a ticking clock. Frontier models rely on web-scale crawls for training, but the open web is rapidly turning into an echo chamber of machine-generated text. Training future architectures on synthetic exhaust risks model collapse and amplified hallucinations.

We think this study confirms what many have quietly suspected: the era of the open, discoverable web as a high-quality commons is winding down. The truly valuable training data and human discourse will increasingly migrate behind paywalls, private discords, and verified credential layers.