The Web Has a Second Audience: Reading Cloudflare's Agent Traffic Data
Cloudflare's September 30 analysis of the agentic web reports that daily requests from AI agents on its network grew by more than 1,700% over a year. The company argues that software acting on behalf of people is becoming a distinct audience for websites.
The measurement needs its denominator. It describes Cloudflare's observed network activity, not the proportion of AI visitors on every website. Automated traffic includes several categories, and request counts are not people, sessions or customers. The growth figure is useful evidence of change without being a universal benchmark for a site's analytics dashboard.
Different requests have different purposes
A search crawler, a training crawler and an assistant fetching a page for a user may all request the same URL. The site owner may want different relationships with each of them. Search discovery can help people reach a publication. A research assistant may summarize it for one task. Bulk collection for training has another purpose entirely.
This makes classification a practical concern. A single “bot traffic” number cannot explain the value or cost of these visits. Teams need to understand who is requesting content, why the operator says it is requesting it, and what evidence supports that classification.
For example, a documentation team may welcome an assistant answering a developer's question while limiting expensive repeated crawls. A specialist publisher may need a licensing arrangement for downstream reuse. Neither decision follows automatically from a large headline traffic number.
Useful content still needs to be readable
Our interpretation is that serving agents well begins with familiar content work. Product facts should be current, headings should explain a page's purpose, and important information should be available as text rather than only embedded in an illustration.
An agent comparing products needs usable specifications and clear regional conditions. A person needs the story and design that help make sense of those facts. Maintaining a consistent content source can support both experiences while avoiding conflicting claims scattered across channels.
Machine access also has an operational cost. Teams should examine repeated downloads, origin workload and the quality of resulting referrals. A request that does not produce a page view can still represent useful attention, but that value should be measured rather than assumed.
Begin with your own traffic
Cloudflare's analysis is a starting point for questions, not a replacement for a site's evidence. Review the operators actually requesting the site, identify expensive patterns and decide which public resources should be easy to retrieve. Keep account access and content-modification permissions separate from public reading.
A short pilot can focus on one journey: a buyer researching a service or a developer seeking an integration answer. Check whether an assistant can find the right public facts and whether a human can verify them on the page. Record mistakes as content or access problems before buying a new layer of technology.
The second audience does not erase the first. It adds another way people can reach information and another set of relationships publishers must manage. The useful response is a more legible, observable website with deliberate access policies.