Cloudflare reports agentic AI traffic surpasses human traffic

Cloudflare's measurement tool, Cloudflare Radar, shows agentic AI bots now account for 57.4% of web requests across the sites it sees, while humans account for 42.6%, according to data shared publicly by CEO Matthew Prince on X and reported by NBC News, CNET, SiliconANGLE and Tom's Hardware. Prince posted the figures and commented, "Welp, that happened faster than I predicted," adding in a follow-up that the underlying data is "a bit messy." He noted he had expected bots to overtake humans only in 2027.

Reporting defines the surge as coming from "agentic" bots, systems that autonomously search the public web on behalf of a user-facing AI assistant, rather than traditional crawlers and indexers, per CNET and SiliconANGLE. Journalistic accounts note these agents can visit far more pages per task than a person: an AI assistant may query hundreds or thousands of pages to answer a single prompt where a human might visit only a handful.

The split is uneven by geography. CNET and SiliconANGLE report North America skewing to about 68.6% agentic traffic, while the US Midwest is an outlier where humans still lead at 54.5% versus 45.5% for agents. SiliconANGLE frames the milestone as lending weight to the "dead internet" theory.

agent-driven requests change traffic shape because they generate high-volume, distributed requests across many domains for a single user intent. This amplifies background request rates and complicates capacity planning, caching, bot classification and rate-limiting. Observers in the coverage note that legacy heuristics for bot detection may not map cleanly onto agentic behaviour, raising observability and cost questions for operators and publishers.

watch for follow-up Cloudflare Radar releases with longer time series and clearer classification methodology, and whether major assistant providers (reporting cites ChatGPT and Gemini as typical initiators) publish telemetry about web queries made on users' behalf. Also watch whether CDNs and publishers adjust caching, attribution and billing, an area Prince has previously tied to "pay to crawl" proposals, as reported by The Decoder.

The figures are self-reported Cloudflare Radar measurements shared via Prince's posts; he and the reporting describe the classification as imperfect and "a bit messy."