Check this page with an assistantOpens a chat asking it to summarise this article and name the evidence behind each claim.
Claude opens with the prompt on your clipboard: Anthropic does not support prefilled prompts on the web, and we would rather copy it than ship a button that drops it.
The Traffic You Cannot See
An agentic browser renders your page in Chromium, reports a Chrome-shaped user agent, and in several implementations arrives without a referrer because the link was opened inside an internal sandbox. The result is a session that looks like a direct visit from an ordinary desktop Chrome user. Standard analytics has no way to tell you otherwise.
This is a different problem from the assistant-referral gap. There, the visitor is a person who clicked a citation and the referrer went missing. Here, the visitor may not be a person at all in the usual sense: software is reading the page, evaluating it against a goal, and possibly acting on it, with a human waiting for a result somewhere else. Both end up in the same Direct row for different reasons, which is worth separating when you diagnose your numbers per the GA4 attribution guide.
The measured picture comes from bot-detection vendors rather than from analytics products, because identifying this traffic requires signals analytics does not collect. HUMAN Security's tracking of agentic traffic through April 2026 put Comet at 48.12% of agent activity, Atlas at 21.33%, the Claude Chrome extension at 17.33%, and ChatGPT Agent at 8.55%, with media at 45.62% of destinations, ecommerce at 38.20%, and travel at 14.12% together accounting for the overwhelming majority.
Why Naming Browsers Is a Trap
OpenAI launched Atlas for macOS in October 2025 and announced its retirement on 9 August 2026, moving agentic browsing into the ChatGPT app, a cloud browser, and a Chrome extension. A product with over a fifth of measured agent traffic lasted under a year. Any advice organised around specific browser names has a short shelf life.
Notice what the retirement does and does not signal. It is not a retreat from agentic browsing; the capability moved into surfaces people already use. The hard part was never the automation, it was persuading people to migrate their bookmarks, passwords, and habits into a new standalone browser. Delivering the same behaviour through an extension or an existing app removes that friction entirely, which suggests the volume goes up rather than down.
So the durable unit of analysis is the behaviour: software rendering your page, interpreting the interface, and attempting a task. That description covers Comet, extension-based agents, cloud browsers, and whatever replaces them next year. Build against it and you do not have to re-plan every time a product is announced or discontinued.
What Makes a Page Operable by an Agent
An agent reads structure, not design. It looks for elements that declare what they are: a button that is a button, an input with a label, a link with text describing its destination. Anything conveyed only through visual arrangement, colour, or position has to be inferred, and inference is where these systems fail.
| Pattern | What an agent sees | Fix |
|---|---|---|
<div onclick> styled as a button | A container, not an action | Use <button> so the role is declared |
| Unlabelled form input | A field of unknown purpose | A real <label> bound to the input |
| "Click here" or an icon-only control | An action with no stated outcome | Text that names what happens |
| Content that appears only after JavaScript runs | Possibly nothing, depending on timing | Server-render what matters, per the JavaScript SEO guide |
| Hover-only menus, drag interactions | An unreachable path | Provide a click or keyboard route to the same place |
| Price or stock rendered as an image | An image | Text plus structured data |
| Interstitials and consent walls | A dead end | Ensure a completable, standards-compliant path |
Read that list again and notice what it is. Every row is a standard accessibility recommendation that predates agentic browsing by more than a decade. Screen readers need declared roles and labelled controls for exactly the same reason: they consume structure rather than layout. This is the unusual case where the AI-era advice has independent justification, an existing standard, and a legal rationale in many jurisdictions.
The Decisions That Are Not Technical
Once software can complete flows on your site, two questions stop being hypothetical: what happens at machine volume, and who is accountable for an action nobody personally took. Both are operations questions, and both arrive before any protocol does.
- Rate-limit anything that costs you something.
Forms that notify a human, actions that consume inventory, anything that sends an email or triggers a callback. Volume assumptions built around human browsing do not survive contact with software that never gets bored.
- Decide what an unauthenticated agent may do.
Browsing and reading, almost certainly. Submitting an enquiry, probably. Anything touching an account or a payment needs the same authorisation you would demand of any other client, and the fact that a browser is driving does not change that.
- Make actions reversible.
Delegated actions are disputed more often than direct ones, because the person did not personally make the choice. Cheap cancellation is the difference between a support ticket and a chargeback, and it is the same conclusion reached in agentic commerce readiness.
- Check your crawler policy is not refusing customers.
A blanket AI block can turn away a user-triggered fetch made on behalf of a real person who wanted to buy from you. Separate the training decision from the traffic decision using the crawler decision guide.
- Watch the Direct segment for shape changes.
Direct sessions to deep pages, unusually fast page sequences, and desktop Chrome sessions with no engagement pattern are the available tells. None is conclusive, which is why this is monitoring rather than measurement.
What Not to Do About It
Do not build a separate experience for agents, do not attempt to detect and serve them different content, and do not block them reflexively. The first is unmaintainable, the second is cloaking, and the third refuses customers who arrived through an unfamiliar door.
The cloaking point deserves emphasis because the temptation is real. Serving different content based on who or what is requesting the page is against Google's spam policies regardless of intent, and the detection you would need is unreliable anyway given that these browsers deliberately present as Chrome. A page that works for everything is both the compliant answer and the cheap one.
The related standards work is worth understanding but not acting on yet: a browser API that lets a site declare what an agent may do, rather than leaving it to interpret your interface, is in origin trial and covered in our WebMCP guide. Until something like that settles, the honest state of play is that agents are guessing at your interface, and your job is to make the guess easy.
Questions People Ask About Agentic Browsers
- What is an agentic browser?
A browser that can operate a website on the user's behalf rather than only displaying it. The person states a goal, and the browser navigates, reads pages, fills forms, and clicks through flows to accomplish it. Perplexity's Comet and OpenAI's Atlas were the best-known standalone examples, alongside extension-based agents that add the same capability to an ordinary Chrome install.
- Why does agentic browser traffic show as Chrome in analytics?
Because these browsers are built on Chromium and report a Chrome-shaped user agent. Analytics tools categorise by that string, so the sessions land in the Chrome row alongside ordinary human browsing. There is no toggle that separates them, which means most sites have this traffic already and no way to see it in a standard report.
- Do agentic browsers pass a referrer?
Often not. Some open links in an internal sandbox or webview that strips the referrer header, so the resulting session is classified as Direct. Combined with the Chrome user agent, that makes agent-driven visits close to invisible in default analytics: they look like a direct visit from a normal desktop browser.
- Is ChatGPT Atlas still available?
OpenAI announced it was retiring Atlas on 9 August 2026, less than a year after launching it for macOS in October 2025, folding agentic browsing into the ChatGPT app, a cloud browser, and a Chrome extension. The capability is not going away; the standalone product is. That churn is the main reason to build for the behaviour rather than for any named browser.
- What should I actually change on my site for agentic browsers?
Almost entirely things you should already be doing: real button and form elements rather than clickable divs, labels on every input, descriptive button text instead of 'click here', a page that works before JavaScript finishes, and flows that do not depend on hover or drag. This is the same work that makes a site usable with a screen reader, which is why it is the rare AI-era recommendation with a decade of independent justification behind it.
