A viral stat says AI agents are now over half of all internet traffic — but does that number actually measure agents?

This explores whether the claim that AI agents now make up more than half of internet traffic holds up, and what that number actually counts. The short answer from the corpus: non-human traffic probably has crossed the halfway mark, but 'AI agents' is the wrong label for most of it. Cloudflare's network data is the main source Is non-human traffic really half of all internet use?. Look closely at the figures, though. The striking 52% describes the share of *crawler* requests made to train AI models, up from 22% in spring 2025. That's mostly automated scrapers collecting text for training. It isn't agents booking flights or shopping for people. 'Bots are half the web' and 'AI agents are half the web' are different claims, and headlines often merge them.

The less obvious point is that this share can rise without agents doing much more. A ratio also moves when its denominator shrinks. Cloudflare reports that human visits to heavily crawled categories fell by as much as 40% in a year. A randomized field experiment found Google's AI Overviews cut clicks to publisher sites by 39.8%, with no gain in click quality or user satisfaction Do AI Overviews actually reduce clicks to publisher sites?. When people get answers without visiting the source, human traffic drops while crawlers keep fetching pages. The non-human share rises even if bot activity stays flat. So 'half the traffic is AI' partly means 'humans stopped clicking.'

It helps to put the 50% figure next to statistics about *agent-driven* activity, meaning AI acting for a person. Those numbers are much smaller. Adobe's widely shared 693.4% growth in AI-referred retail traffic was real. But Lipsman shows it started from almost nothing: AI referrals went from 0.1% to 0.6% of ecommerce traffic Does Adobe's AI referral traffic growth statistic hold up under scrutiny?. That's a useful habit for reading any AI-traffic headline: ask about the baseline, and ask whose behavior is being counted. A huge growth rate and a majority share can both be true and still tell you little about how much agents actually do.

Other notes suggest the internet is becoming machine-shaped in more ways than traffic. On the supply side, about 35% of newly published websites by mid-2025 were AI-generated or AI-assisted, and the range of topics and meanings across new pages narrowed How much of the internet is AI-generated now?. On the people side, generative AI adoption reached 54.6% of US adults within three years, faster than the PC or the internet spread Is generative AI adoption now outpacing the personal computer?. Put together: machines increasingly write the pages, machines read them, and humans see the results through AI summaries. The 50% traffic figure may matter less as a count of agents than as a sign that the old setup, where people clicked through to sites and publishers earned money from those visits, is coming apart. That's why Cloudflare notes publishers are turning to licensing deals instead of relying on search referrals.

One gap: the corpus has no independent measurement that separates training crawlers, search indexers, and true user-directed agents in overall web traffic. The 'agents' part of the claim is the least supported.