AI Agent Traffic

agent traffic · agentic traffic · AI agent visits · agent-driven traffic

Definition

AI agent traffic is visits to a website that an AI agent makes on a person's instruction in real time, as opposed to background crawling of the site for model training or indexing.

What AI agent traffic is

AI agent traffic is visits driven by a specific person’s specific task, carried out by software. A shopper asks an assistant to find running shoes under $150 with next-day delivery — the agent opens several stores’ catalogues, compares product pages and, if it has permission, places the order. In July 2026 Cloudflare defined this kind of traffic as automated behaviour acting on a person’s behalf, usually in real time, to get something done right now.

Agent, crawler or scraper

Visit type	Why it comes	Examples	robots.txt
Agent on assignment	Carries out a person’s task right now	ChatGPT-User, Claude-User, Perplexity-User, Google-Agent	OpenAI, Perplexity and Google say their fetchers may not follow it; Anthropic says Claude-User does
AI search crawler	Indexes pages for assistant answers	OAI-SearchBot, Claude-SearchBot, PerplexityBot	Follows it
Training crawler	Collects data for models	GPTBot, ClaudeBot	Follows it
Scraper	Collects prices and content for third parties	Disguises itself as a browser	Ignores it

The second and third types are covered in more detail under AI crawlers.

How to recognise an agent

Declared User-Agent. AI companies’ user-triggered fetchers identify themselves by name, and OpenAI and Perplexity publish lists of their IP addresses for verification.
Signature. Agents that sign their requests with Web Bot Auth can be verified cryptographically — the most reliable signal.
Behaviour. Dozens of product pages a minute, navigation with no scrolling or mouse movement, forms filled in within fractions of a second.
Agentic browsers. An agent working inside the user’s agentic browser may be indistinguishable from a person by both User-Agent and IP.

How it distorts metrics

Metric	What happens without segmentation
Conversion rate	An agent views many product pages across several stores and buys in one — the others’ CR drops
Bounce rate	Short open-read-leave visits look like bounces
Unique visitors	An agent in a clean browser with no cookies looks like a new visitor on every run
A/B tests	Noise in the metrics and a risk of SRM
Recommendations	Agent views leak into the popular and viewed-together behavioural signals

What a store should do

Measure the share of agents from server or CDN logs, not only in analytics: analytics filters catch known bots, but not every agent.
Put agents into a separate segment and a separate funnel: view, basket, order.
Exclude them from A/B tests and from the data used to train recommendations.
Do not confuse them with AI referral traffic — people who clicked a link in an assistant’s answer.
Check your CDN settings: from 15 September 2026 Cloudflare blocks the Agent and Training categories by default on pages with ads for newly onboarded domains.

Frequently asked questions

How is agent traffic different from ordinary bots?

Ordinary bots — search crawlers, monitoring tools, scrapers — run on their own schedule and serve their operator's interests. An agent arrives because a person has just given it a task: find, compare, buy. There is real demand behind the visit, so blocking it the way you block a scraper means losing customers.

Can GA4 and other analytics tools see agents?

Only partly. GA4 automatically excludes traffic from known bots, based on Google's research and the IAB list, and you can neither switch the exclusion off nor see how much was excluded. An agent working through an ordinary browser can look like a person and slip past any such filter. Check how your analytics platform treats bots detected by behaviour: some, such as Yandex Metrica, keep them in reports by default until a filter is switched on.

How does agent traffic affect A/B tests?

Agents do not respond to design and copy the way people do, so they dilute the effect of the variants. If agents land in the variants unevenly — for example, one variant breaks their flow and they leave earlier — you get a sample ratio mismatch (SRM), and the test result cannot be trusted. It is better to exclude agents from experiments or analyse them as a separate slice.

How much agent traffic do websites get today?

There is no single measurement; the share depends on the niche and on the site's protection. CDNs give general reference points: according to Cloudflare, as of July 2026 more than half of internet traffic is no longer human, but that covers all bots, not only agents. In Cloudflare's report for the 12 months to July 2025, user-requested actions accounted for about 2% of AI bot crawling, training for 80% and search for 18%. The most reliable way to measure the share on your own site is from server logs.