An agentic browser is a browser that ships with an AI agent capable of acting on the page, not just rendering it. The user states a goal — "find me a cheaper plan," "book the earliest flight," "compare these three vendors" — and the agent navigates, reads, clicks, and fills forms to accomplish it. ChatGPT Atlas and Perplexity Comet are the best-known examples.
For anyone measuring or optimizing a website, agentic browsers create a problem that ordinary bots do not: the traffic is human-initiated but machine-executed. A real person wants something, so the visit is not fraudulent or worthless. But a model is doing the reading, so nothing you optimized for human perception — hero imagery, visual hierarchy, social proof placement, above-the-fold layout — reaches the entity making the decision.
Why They Are Hard to Detect
Agentic browser traffic largely defeats conventional bot filtering:
- The user agent looks like Chrome. ChatGPT Atlas presents a user agent string matching stock Chrome. A user-agent filter cannot separate it from an ordinary visitor.
- It runs real browser infrastructure. JavaScript executes, cookies are set, and viewport dimensions look plausible, so heuristics built for headless scrapers do not fire.
- There is no dedicated robots.txt token for ordinary agentic browsing, so you cannot express a policy about it the way you can for a declared crawler.
The most reliable identification method available is cryptographic rather than heuristic: OpenAI's ChatGPT agent signs its requests using HTTP Message Signatures (RFC 9421), carrying a Signature-Agent header that can be verified against published keys. That covers declared agent activity — it does not cover every agentic session.
Agentic Browser vs. Crawler
| Crawler | Agentic browser | |
|---|---|---|
| Initiated by | A schedule or index job | A specific human request |
| Declares itself | Usually, via user agent | Often not |
| Executes JavaScript | Frequently not | Yes |
| Converts | Never | Sometimes — on the user's behalf |
| Should you block it | Depends on policy | Usually not; it represents a real customer |
Why It Matters for Experiments
Agentic sessions get bucketed into experiment variants like any other traffic, but they rarely convert on the metric you are measuring. That dilutes observed conversion rates, inflates variance, and lengthens time to significance. If agent traffic reaches variants unevenly — which redirect-based tests make likely — it can also trigger sample ratio mismatch and invalidate the test outright.