If someone had asked me a year ago what might be the next big innovation to come out of the major AI companies, I’m pretty sure I wouldn’t have said, “a browser.”
But that’s exactly what both OpenAI and Perplexity did, each launching their own shiny new AI-enabled browsers, Atlas and Comet, respectively.
If you read the PR comms or watch the launch demos, both companies frame their new browsers as the first step towards completely reshaping how regular consumers use the internet. In OpenAI’s livestream to launch Atlas, Sam Altman said that “AI represents a rare once-a-decade opportunity to rethink what a browser can be about.”
Over on Substack, OpenAI’s CEO of Applications, Fidji Simo, wrote about Atlas and ChatGPT “evolving to become the operating system for your life.”
Perplexity CEO Aravind Srinivas has used similar language, describing Comet as a “cognitive operating system,” while Jesse Dwyer, who worked on Comet, is widely quoted referring to the browser as the “operating system of your mind.”
This all sounds extremely transformative. But I just don’t see it. At least not yet.
While these phrases might have polled well in focus groups, they’re effectively meaningless. A browser is not an operating system in exactly the same way ChatGPT or Perplexity aren’t.
Chrome, Edge, and Safari are each tied into a huge suite of digital products and tools centered around a different operating system. While it’s true that anyone can install and use any of these browsers, it’s that deep integration with a comprehensive suite of proprietary tools that creates ecosystems and builds workflows.
Could it be that, in their mad scramble to find sustainable monetization models, OpenAI and Perplexity have both fallen victim to survivorship bias?
What Tech Companies Can Learn From World War II Bombers
During World War II, the U.S. military tasked the Statistical Research Group at Columbia University with solving a pressing problem: too many American bombers weren’t returning from missions over Europe. Examinations of the aircraft that did make it back revealed patterns in the damage, with bullet holes heavily concentrated on the fuselage and wings. The obvious conclusion appeared to be to reinforce those heavily damaged areas.
But mathematician Abraham Wald saw the problem differently. He realized the military was only looking at planes that survived. No matter how shot up a plane might be, it could only make it back to base if none of those hits were critical. What about the planes that didn’t come back?
Far from exposing weaknesses to be reinforced, the bullet holes revealed where damage was survivable. They certainly did not reveal which areas were decisive in determining a bomber’s success or survival.
This is possibly the most famous example of survivorship bias: where we mistakenly focus on the common traits of those that succeeded (or survived) while ignoring the many others that failed, leading to false conclusions about which aspects genuinely contributed to that success.
Survivorship bias is all around us. Articles cite university dropouts like Bill Gates and Mark Zuckerberg as proof that degrees don’t matter, while conveniently ignoring the vast number of dropouts who aren’t billionaires. Countless self-help books promise to unlock the formula to fame and fortune by documenting the habits of successful people – waking at 5 a.m., meditating daily, taking cold showers – while ignoring the millions who follow similar routines and never make it.
In a similar vein, Perplexity and OpenAI appear to have looked at the successful tech giants they hope to one day emulate – Google, Microsoft, and Apple – and decided a common factor to their continued success is that they each have a proprietary browser.
When Imitation Looks Like Insight, But Isn’t
Last year, when it briefly appeared that regulators might impose structural remedies on Google, including divestment, both OpenAI and Perplexity expressed serious interest in buying the Chrome browser. Perplexity even went so far as to put in an unsolicited bid for $34.5 billion.
Unfortunately for them, while the judge did impose penalties and other restrictions on Google, he did not impose a forced sale of Chrome.
So, it shouldn’t be surprising that both Atlas and Comet look extremely familiar, particularly as they were both in development while Google’s anti-trust case played out in court. This is Plan B. If you can’t buy the browser, copy it.
OpenAI and Perplexity looked at Chrome’s well-established user experience – with its unified address bar, tabs, extensions, and more – and replicated all of it. Both browsers are even built on Chromium, the open-source technology that drives Chrome.
Then again, so did Microsoft when building Edge. And while Safari and Firefox aren’t built on Chromium, they have adopted many of the features first popularized by Chrome.
Of course, both OpenAI and Perplexity have each layered their own AI secret sauce over the top of the usual bells and whistles. Both browsers are designed to support “agentic browsing,” handling all that apparently tedious clicking around and reading stuff on your behalf. (Call me old-fashioned, but I like surfing the web. I like browsing and discovering and stumbling across things. I like reading a well-written article far more than a short summary of the key points.)
Agentic browsing can even find and book holidays, manage your email, and complete your shopping, all while you do something else.
The thing is, we already have agentic AI. A lot of what these new browsers can do already happens when you use ChatGPT or Perplexity anyway. It’s just that it previously happened behind the scenes.
When you type in a query or prompt, the LLM uses its headless browser to search the web and find the information required to generate an accurate response. These new agentic browsers make that previously hidden process visible. You can watch as the browser, say, restructures the data in a spreadsheet, or jumps from page to page on a supermarket website, adding items to your basket ready to check out.
Or, more likely, you’ll probably do something else in another tab or window until the browser notifies you that the task is complete, because who needs to see the sausage being made?
Measurement, Fraud, And Security Headaches
The AI companies clearly hope the promise of agentic browsing will entice people over to Atlas or Comet. However, it’s likely to create a bunch of headaches for organizations.
While an LLM’s headless browser might identify itself as, for example, PerplexityBot when visiting your website, the same isn’t true when Perplexity’s Comet browser visits your site on a user’s behalf. To your analytics tool, it probably looks like any other Chromium browser, complete with the visitor’s IP address. In other words, you probably won’t be able to tell whether it’s a human or an agentic browser visiting your site.
As Digiday points out, this creates a whole heap of problems for marketers and SEOs. First, your usual metrics around traffic and clicks become less reliable. If you can’t measure it, you can’t manage it.
Then there’s the potential for massive advertising fraud. If ads are served to AI agents instead of humans, who’s to know? And with AI agents capable of thousands of requests per second, things could get real scary.
And then there are the security implications.
These new agentic browsers introduce significant new security risks. Gartner has recommended that companies block all AI browsers over cybersecurity concerns, specifically citing prompt injection as a major risk.
Last year, LayerX Security coined the term CometJacking to describe one such technique. Simply by clicking a malicious link, the user triggers hidden commands that instruct Comet’s AI to access and steal any sensitive data exposed in the browser. The hackers don’t need to phish for passwords or any other credentials because the browser already has authorized access to everything.
Just three weeks later, LayerX discovered a vulnerability in Atlas that could also be exploited by bad actors to inject malicious instructions into ChatGPT’s memory.
Right now, instructing one of these browsers to book tickets, complete transactions, or edit important documents on my behalf feels less like getting help from a trusted assistant and more like handing your keys to a complete stranger.
What Problem Do These New Browsers Actually Solve?
But that doesn’t mean these new agentic browsers won’t have a place. While mainstream adoption looks unlikely, some digital teams could still find real value working with these agentic browsers.
Never mind laborious manual testing. Developers and UX testers could use these browsers to simulate user journeys at scale, testing how websites respond in different scenarios far more efficiently.
SEO professionals might use them to understand how AI agents interpret site structure and webpages, providing clues to how the LLM’s hidden headless browser “sees” your content and highlighting where improvements might be made to improve visibility.
And for technical users comfortable with the security risks, there’s no doubt that agentic browsers offer genuinely useful automation for repetitive tasks, such as extracting data from multiple sources into spreadsheets or monitoring specific websites for changes.
The irony is, of course, that these power users and developers are completely the wrong audience if OpenAI and Perplexity are hoping to capture oodles of valuable user data to further train their models. They need data on typical consumer behavior patterns, not developer testing workflows.
If Atlas and Comet are to have any chance of differentiating themselves from the long-standing incumbents of Chrome, Safari, Edge, Firefox, et al, they need to offer a clear, attractive, and significant value proposition to average consumers; ideally, one that isn’t fraught with risk. But as the major browsers have also begun to gradually integrate with AI, that proposition is far from clear, and any differentiation is likely to be short-lived.
If AI is going to transform how most people browse and interact with the internet, it won’t be because of Atlas or Comet. More likely, it’ll be because of Chrome or possibly Edge. Not only do these browsers already have the juice as longstanding incumbents, but both Google and Microsoft also have their own extremely powerful proprietary LLMs.
Firefox could be worth watching too, as it launches new AI controls designed to give users far greater control over which AI features they want to use, or block. Firefox’s approach might give us the clearest picture yet of the kinds of AI-powered experiences users really want to see, instead of what AI companies would prefer.
The “AI browser war” has barely begun, and I don’t think it’ll be short. This isn’t going to be a rapid disruption, no matter how much OpenAI and Perplexity would like to somehow skip decades of incumbency and trust with a few flashy demos and a bunch of optimistic predictions. Instead, the winners will be those who focus on the user experience and that all-important value proposition above everything else. Whatever form AI or agentic browsing eventually takes, we’ll know the war is over when it disappears so completely into people’s workflows that nobody even thinks about them anymore.
More Resources:
- Marketing To AI Agents Is The Future – Research Shows Why
- Why Agentic AI May Flatten Brand Differentiators
- Explaining Agentic SEO To The C-Level
Featured Image: Who is Danny/Shutterstock
