OpenAI's Atlas To Perplexity's Comet: AI Browsers Aren't Smart Enough Yet To Take Over Internet

When OpenAI unveiled an artificial intelligence-based browser in October, Alphabet Inc. investors expressed concern about what it would mean for Google Chrome, the ubiquitous gateway to the internet used by billions globally.

And yet, current versions of AI browsers are far from making legacy products like Chrome obsolete: New offerings from companies such as OpenAI and Perplexity AI Inc. present occasional bugs and trip over some seemingly straightforward requests, Bloomberg News found after testing them for a month. 

The new browsers, including OpenAI’s Atlas and Comet from Perplexity, put artificial intelligence assistants front and center, replacing search engines as the preset option when users enter requests. Many also offer a feature called agentic browsing, enabling them to carry out multi-step tasks on behalf of users, like completing shopping orders and extracting a list of to-dos from unread emails. 

The goal, as AI developers see it, is for consumers to use their chatbots not just from their own apps or websites, but within browsers and mobile operating systems, potentially opening more avenues for ad targeting and revenue streams. The most advanced features are currently only available on a paid tier, given that AI agent features can be more expensive to run.

For now, the two categories of browsers are encouraging different kinds of user behavior, forcing app developers, web services and publishers to rethink whether they’re designing their tools for humans or robots crawling the web — and whether they can be served by the same products. 

“We are seeing folks change the way that they’re interacting with the web and how they’re finding information in different categories,” said Laura Chambers, chief executive officer of Mozilla Corp., which develops Firefox.

Legacy developers like Google, Microsoft Corp. and Mozilla have taken a different approach, keeping search as a default in their browsers, while adding AI assistant features over the past year.

“People are still using Google search for things like shopping and travel, areas where you really want to research, look at a bunch of different things and make purchasing decisions,” Chambers added. 

A Mozilla user research survey conducted last year found that 60% of participants were only comfortable using generative AI for low-stakes matters or things they know enough about that they can readily verify the quality of output and spot mistakes. The results, which haven’t been previously reported, were based on a survey of 1,000 people in the US and 400 Firefox users in May 2024. 

Chambers said generative AI is more helpful with informational searches because it can synthesize information and provide assorted links that users can refer back to. The same survey found that saving time can encourage users to replace search with generative AI, but a need for accuracy drives them to combine the two methods.

In user research interviews, browser developers have hailed productivity wins by their AI-assisted products and observed a desire from consumers for web browsers that can do more. 

A popular use for these AI browsers has been summarizing long YouTube videos and asking the assistant about a topic mentioned in the video, according to representatives from OpenAI and Chrome. Comet users are also asking “six to 18 times more questions” than they did with the regular Perplexity chatbot, said Jesse Dwyer, the company’s head of communications.

And consumers are wanting more. Adam Fry, a product lead working on OpenAI’s Atlas browser, said OpenAI has been getting a lot of requests from power users for the ability to schedule tasks so that the browser can repeat them on a regular basis. 

For example, a finance professional might employ such a feature to generate a report or dashboard from an online tool every month. Ted Choc, a director of engineering for Chrome, said users want the browser to fill out government forms on their behalf and do their taxes.

Some of these capabilities were technically possible even before the era of large language models, though those tools weren’t as easily accessible. Legacy browsers have been able to auto-fill forms if you have pre-saved your address, credit card information or other personal details. YouTube and third-party tools can transcribe videos. Programmers can write computer scripts to carry out scheduled tasks in bulk. 

But there hasn’t been an urgent reason for desktop browsers to evolve, especially since mobile took over desktop as the main source of web traffic. According to Similarweb, more than 67% of it today comes from smartphones and tablets, where mobile operating systems and app stores serve as a gateway to standalone apps.

Now, AI chatbots that take instructions in natural language are letting users ask more of their devices to get things done.

Ben Goodger, Atlas’s head of engineering and a Google veteran who helped create Chrome, said on a company podcast last month that AI browsers are helping “people think less about the particulars about the tools they are using, more just expressing what they want to the system.”

But the way websites are built today represents a significant obstacle. Artificial intelligence players like Anthropic PBC and OpenAI have developed new protocols so that digital service providers and AI agents share a common language.

How well AI assistants provide answers and complete tasks will depend on how websites are structured. Websites that are heavy on visual elements or are narrative-oriented are not necessarily useful, said Linda Tong, CEO of Webflow, a web development and analytics tool. 

“Agents want very structured, well-defined indexable data and insight that is pretty specific,” Tong said. “I do believe that there is going to be a world where you need two versions of your site, and we’re already starting to make that shift. You're going to actually want to share very different information on a human version of a site.”

A “robot” version of websites would improve browsers’ ability to act as AI agents, the key feature separating AI browsers from regular ones. It essentially enables the browser to execute mouse clicks, cursor placements and keyboard input based on a user’s instructions to the chatbot. The idea is that the chatbot can more easily execute tasks if the website is structured behind the scenes in a way that it can easily understand, rather than having fancy buttons or cards that appeal to humans.

Most of the web today is still built for humans, causing some of the AI browsers’ more advanced features to stumble. Some of these symptoms are also a result of what OpenAI’s Fry described as Atlas browser being “too smart for its own good” — meaning, it overthinks some simpler tasks like clicking and editing documents.

When a Bloomberg reporter asked the Atlas browser to scroll through the reactions to a LinkedIn post and make a list of those working in a particular industry, it took a minute or two to decide how it would do so. After that long pause, it opted to take a screenshot of each section while it scrolled, performing image recognition so it could interpret the text on screen. 

In another instance, it became stuck in a loop, unable to complete the instructions. Perplexity’s browser, meanwhile, almost always defaults to screenshotting when it encounters more complex visual components like pop-ups.

In Bloomberg’s tests on a Windows laptop, the machine heated up and the fan got louder when some of these advanced agent features were in use. While the problem of browsers hogging computing resources is nothing new, legacy browser developers have had a head start in trying to address these problems.

The AI browsers fare better with simpler tasks, like summarizing contents on a screen, or searching for information online and entering what it found into online text editors and spreadsheets.

The landscape has shifted even since Atlas debuted in October, or at least the optics have. Google’s latest Gemini 3 model outperforms OpenAI’s best AI systems on many benchmarks, which prompted CEO Sam Altman to declare a “code red” to improve its flagship chatbot. The browser playing field is meanwhile set to level as Chrome and Microsoft bring AI agent tools out of the testing phase for broader release within their legacy browsers.

What’s more, getting online providers to develop AI-specific versions of their services has been a largely thankless endeavor. Just ask OpenAI: The company released a framework this fall for app developers to integrate their services into its chatbot, and isn’t always getting the buy-in it wants. While services like Zillow, Instacart and Booking.com have piloted versions of their apps within ChatGPT, their capabilities are limited enough that users would be better served using native websites or apps. Some developers, such as Airbnb, say they’re not rushing to get on board.

Other firms operating high-traffic, including Reddit Inc., Amazon.com Inc. and the New York Times Co., have filed lawsuits objecting to how their content is being leveraged by AI companies. In its suit against Perplexity, the New York Times specifically called out Comet’s ability to provide summaries on articles that should only be available to the paper’s paying subscribers. 

Perplexity has vowed to push back against what it sees as bullying tactics from legacy providers that are threatened by AI upstarts.

For users, the AI browsers’ under-baked nature could pose a barrier to wider adoption. 

“AI is not yet trustworthy, and if we want to get to a point where you’re depending on an agent to spend money on your behalf, we're going to have to really, really increase that trustworthiness,” said Mozilla’s Chambers. 

“If I'm booking a flight, how do I know that that agent is really giving me the best deal for me rather than getting a kickback for booking me on airline number X?” she said. “One of the big risks with AI is that it's going to be deeply integrated into the experiences. And so people are very worried about that.”