In October 2025, OpenAI launched ChatGPT Atlas, a full web browser built around an AI agent that could read pages, fill forms, and click through websites on your behalf. On August 9, 2026, less than ten months later, OpenAI shut it down and folded its browsing features into ChatGPT and Codex instead. That is not a footnote. It is a sign that even the company betting billions on AI agents is still figuring out whether a standalone AI browser is the right product, at the same time that AI is already reshaping how people find information online. Those two threads, product churn and traffic collapse, are worth pulling apart separately, because they tell different stories.
Atlas lasted 292 days
OpenAI has not called Atlas a failure outright, but the timeline speaks for itself. The browser crossed 5 million monthly active users earlier in 2026, which is a real number for a new product, but OpenAI decided the agentic browsing features worked better living inside ChatGPT and its coding tool Codex than inside a separate browser people had to switch to. Users who wanted to keep their Atlas bookmarks, tabs, and history had to manually export them, since none of it carried over automatically.
Perplexity took a different bet with its own AI browser, Comet, released in mid-2025 and made free to everyone by October 2025. Comet has reportedly reached around 3 million monthly active users. The difference matters: Perplexity is trying to own the browser as the front door to the internet, while OpenAI just concluded that a separate browser was not the right wrapper for the same underlying idea. Nobody has proven yet which approach wins, which is the honest, unresolved part of this story.
Two AI browsers, two different footprints
There is a smaller but genuinely interesting technical wrinkle here. When Comet visits a website on a user's behalf, it typically shows up in that site's analytics as a normal referral from perplexity.ai. Atlas, by contrast, often stripped or blocked referrer information, so traffic it generated tended to show up as "Direct" or simply unlabeled in tools like Google Analytics. That meant two AI browsers built for the same basic task could look completely different in a website owner's traffic reports, one visible, one invisible, even though both were sending real visitors.
The bigger number: traffic is actually falling
Separate from any single browser, the broader trend for website owners is not subtle. Analytics firm Chartbeat data covering more than 2,500 publisher sites found that Google search referral traffic fell roughly a third over the course of 2025, with a steeper 38 percent drop in the United States specifically. Google Discover referrals dropped by close to 30 percent over the same period. Search Engine Land has separately reported that pages where Google AI Overviews appear saw organic click-through rates fall by 61 percent.
A widely cited industry report from TollBit, a company that tracks AI bot activity on publisher sites, found that AI chatbots such as ChatGPT and Perplexity send 95 to 96 percent less referral traffic than a traditional Google search does, with an overall referral rate of just 0.37 percent. Put simply: AI systems are reading a huge volume of web content and answering questions directly, without sending nearly as many actual readers back to the source.
- National and local news sites saw roughly 70 to 76 AI bot visits for every one human visitor, according to TollBit's data.
- Deals and shopping content saw an even more extreme ratio, over 2,000 bot scrapes per human visitor.
- Despite the low volume, AI-referred traffic that does convert can do so at rates close to organic search in some categories, TollBit found B2B software conversion rates from AI chatbots and Google organic search were nearly identical.
Why this is not a simple villain story
It is tempting to frame this as AI companies quietly draining value from the websites they depend on for information, and there is truth in that framing. But the picture is messier than a single number suggests. Some AI-referred visitors convert well, meaning the traffic that does arrive can be more valuable per visit even though there is far less of it. And zero-click behavior was already rising before generative AI became mainstream, users have been getting answers without clicking through for years through featured snippets and instant answers.
What is new is the scale and the speed. A jump in AI agent and bot traffic since mid-2025 has been large enough that some site operators now say bot visits to certain pages outnumber human ones by a wide margin, which creates real hosting and infrastructure costs even when no human ever sees the page.
What this means if you run a website or small business
If your business depends on people finding you through search, the practical takeaway is not to panic but to stop assuming Google referral traffic will hold steady. Diversifying how people discover you, through email lists, direct visits, and communities you actually control, matters more now than it did two years ago. It is also worth checking your own analytics for unusually high direct traffic or unexplained spikes in server load, since some of that may be AI agents and crawlers rather than confusion in your tracking setup.
The takeaway
The Atlas shutdown and the traffic decline data are really the same story told from two angles: the tools built to browse the web on our behalf are still being reshaped in real time, and the web they are browsing is quietly changing shape underneath them. Nobody, including OpenAI, has landed on a settled answer for what an AI browser should even be yet. Meanwhile the traffic numbers are not waiting for that answer to arrive. If you build anything online, the safest assumption right now is that the rules of discovery are mid-rewrite, not finished changing.
