I tested five AI browsers and lost my mind in the process

All I wanted was a pair of New Balances. I was done trusting stylish influencers who swore Vans, Converse, and Allbirds were up to the challenge of walking 20,000 steps day in and day out. They are not. Fall is the season of holiday sales, so there’s no better time to shop… if you’re immune to being overwhelmed by modern day e-commerce.

Wouldn’t it be grand if I could skip all the fake deals and barely disguised ads, and have the internet find the best stuff for me? What if I could tell the internet my wish and have it granted?

Tech CEOs have been evangelizing that this is the future. Tell the bot what you want, kick up your feet, and let AI do the rest. Microsoft CEO Satya Nadella recently said on a podcast that one day, AI will be able to “use a computer as well as a human.” He’s far from the only executive touting that bots and agents might soon be better than we are at just about everything.

In the past few weeks, it’s become clear that browsers are the latest entrant in the AI arms race. We’re talking about things like Perplexity’s Comet, ChatGPT Atlas, and even Chrome — browsers that natively embed chatbots into the internet experience. The pitch is to reorient how we browse, to move us away from the search engines that have reigned for the past three decades. The central idea is the same as we’ve heard from all the other agents-all-the-way-down companies: AI will be just as good as you are at surfing the web. Possibly better.

Right now, AI browsers come in two main flavors. There are your regular browsers that have an AI assistant stapled on in a collapsible window, such as Chrome with its Gemini features, or Edge with Copilot Mode. Then there are more specialized AI browsers, most notably ChatGPT Atlas, Perplexity’s Comet, and The Browser Company’s Dia. This second category often supplants your search bar with AI and sometimes includes an “agentic mode,” in which the AI can complete more complex, browser-related tasks for you. Theoretically, that includes helping you book reservations or add items to a shopping cart.

For testing, I decided on a few ground rules. I kept it to five browsers: Chrome, Edge, Atlas, Comet, and Dia. There are more available, but this felt like a representative mix of both AI browser categories from a variety of players in the field. I focused on desktop apps, and tried to make settings as uniform as possible: I generally instructed the AI browsers to keep answers snappy, shared my location information where possible, enabled memory settings, and described myself as a “tech journalist specializing in health and wearable tech.” I also approached testing from a variety of AI skill levels. What would results look like if I was a complete AI newbie versus someone more adept at prompting? Lastly, if I tried one task in a browser, I gave it a go in all the browsers, down to the same exact prompt.

Ultimately, my question was not which AI browser you should use, but whether any of them are worth your time and energy. This was a journey to see whether any of them live up to the hype.

The short answer: they don’t.

Stapling an AI assistant to a browser doesn’t magically redefine how you interact with a chatbot. It’s more like hanging out in person, rather than texting. You’re having the same conversation, just in different formats, each with their own pros and cons. But no matter the browser, I kept running into the same fundamental problem: you have to think extra hard about how to craft the right prompt.

That’s the opposite of how search, particularly Google, has evolved. At Google’s peak, you could type in a string of misspelled words into the search bar, and somehow you’d still get the right answer. AI models require a lot more prep and guidance.

Take the universal torture of sorting your emails. On any given day, I want to know what my most important emails are, and which ones I need to respond to ASAP. The first few times I took a crack at this task, I asked the various browsers to summarize my emails. (I know “summarize my emails” isn’t a stellar prompt, but it’s often a default suggestion. And defaults exist because they should be generally helpful.) All I got were literal descriptions of the emails in my inbox. In my personal inbox, it said I had one thread in my primary folder, listed the subject, summarized the preview, and then stated it was “dated Nov 20th, and was marked not starred or important.”

I tried refining my ask. Instead of “summarize,” I prompted AI to “identify important emails based on urgency.” In my work inbox, that generated a list of not-important, not-urgent email threads because the models have no idea what I actually find important. I wanted reader feedback, pitches from trusted contacts, or threads I’d forgotten to reply to. What I got instead were irrelevant pitches, mostly for health scams.

I made zero headway until Comet suggested the prompt “find important unanswered emails.” The top four emails that surfaced were littered with important keywords for a tech journalist — Urgent! Embargo! Exclusive for The Verge! All had multiple follow-up requests. You can see why Comet would think they mattered, but after reviewing them, all were emails I didn’t need to read at all, much less reply to. AI had fallen for the oldest trick in the book: conflating keywords with truth.

I was ready to write off the experiment when I noticed Comet’s AI had buried the lede. Nestled three-quarters of the way into its long-winded summary was a bullet point labeled “Personalized requests/follow-ups.” It highlighted two emails: one from a CEO addressing feedback I’d made in a recent product write-up, and another from a reader with a tip relevant to my beat. Neither was “urgent” but both merited a closer look.

I tried Comet’s “find important unanswered emails” prompt in the other AI browsers. They all highlighted other previously skipped, keyword-stuffed pitches. None flagged the two emails I was interested in. So I tried even harder:

Find unanswered emails in which I had previously responded with interest or feature personalized requests/feedback. Then, evaluate which ones I should respond to based on timeliness and keywords such as “embargo” featuring dates in the next two weeks. Ignore emails with multiple follow-ups to which I have not responded.

This went slightly better with Comet and Dia. Both surfaced multiple relevant email threads, but only one ultimately required a response. Copilot in Edge highlighted one relevant thread and five junk pitches. Gemini in Chrome was a dud: It surfaced only a Black Friday marketing email.

In Atlas, ChatGPT merely replied, “It looks like Gmail successfully returned the unread message IDs, but the actual content for those messages didn’t come back — the batch read returned empty, which means the Gmail API didn’t provide the email bodies this round.” It proceeded to ask two long-winded follow-ups. At this point, my options were to refine my prompt further or give up.

Emails were mostly a failure, but there were some daily tasks where AI browsers were alright. I had to search through a 48-page legal document for a family matter, and while CMD-F is tried and true, the legalese made my brain melt. So I loaded the document in a tab and prompted the AI browsers to list all the relevant pages and sections, with an accompanying summary. All the browsers returned the same pages, with slightly different summaries. I still had to do my own reading, but it got me to a useful jumping-off point faster.

These browsers can also work well for internet search — provided you’re patient enough to reprogram 20 years of Google muscle memory. Where AI search works best is answering questions about the exact site you’re on. While pondering a phone upgrade, I asked the bots to compile various iPhone specs and size dimensions into a table from Apple’s website and the wider web. That was much more convenient and helpful than flipping through multiple tabs. At the end of the process, I was much more confident about which iPhone I was upgrading to.

Whenever I went into a task asking AI to do things for me, I’d end up frustrated. I was much more successful whenever I shifted my mindset to “how can AI help me interact with this page?” For example, I was trying to parse a clinical study, and hit a particularly technical paragraph written in dense medical jargon. Asking the models to summarize and explain some concepts I was iffy on in plain English was helpful.

Summarizing or compiling data like this was the most convenient part of using AI browsers. All the browsers do this fairly well, and it’s a useful thing to have at your fingertips — it’s not without the occasional back-and-forth, but overall, I needed less time and fewer tabs to get to a point where I’d take over the heavy lifting of getting something done online. I’m always in favor of fewer tabs while browsing.

We already know that AI is good at summarizing and compiling, though. Complex queries are where these browsers are meant to shine. Here, too, one must grease both elbows and wrestle AI into submission.

Ahead of the Stranger Things season 5 premiere, I was chatting with a colleague about watching an 18-minute YouTube recap video. They were separately working on another AI project and asked if AI browser assistants could turn videos into downloadable, .txt transcript files. So I tried prompting: can you rip a transcript of this YouTube video?

Copilot said no, on account of the video’s copyright. (Never mind that most YouTube videos already have transcripts, right there on the page. This should not be a hard problem.) What I could get was a summary or an outline of the video’s content. Comet ripped an accurate transcript for the first 25 seconds before stating that the “transcript continues for Seasons 1-4 with detailed plot and character recaps.” Dia gave a time-stamped transcript, but only for the first 15 minutes. Atlas and Chrome were the only two to give full transcripts. As in, an extremely long, line-by-line transcript right into the chat window.

Next, I asked each browser’s AI if they could turn that transcript into a downloadable .txt file with timestamps. Only Atlas completed the task. The rest said generating a downloadable file wasn’t in the cards, but I could copy-paste the plain text into a file myself.

So much for “just telling the AI what I want.”

After several detours, I returned to my original task: figuring out which pair of New Balances to buy, and finding the best deal possible.

When I say I wanted a pair of New Balances, it’s because I’ve spent about three months researching. I look at social media, I ask friends, I read up on the history of various brands before ultimately choosing one. Then I’ll spend a few hours on that brand’s website whittling down my options until I have about three. Afterwards, I’ll try to find a deal online. It’s a long, arduous process prone to human error. Hence, why it’s been two years since I set out to find a pair of durable, stylish, and comfy walking shoes and I’ve yet to find one.

With AI browsers, the research part was “easy.” In a nutshell, I had to give it highly specific research prompts. That meant telling it that I’m: flat-footed, more comfortable in wide shoes, looking for a lifestyle sneaker and therefore no running shoes, looking for something that can easily handle 15,000 to 20,000 steps per day, interested in a versatile color but prefer a neutral white, wanting something that works with athleisure and elevated street wear, and not looking to spend more than $120 (but would prefer to stay under $100).

What ensued were multiple back-and-forths where the browsers both did and didn’t listen to my needs. The longer the response, the more likely I’d get contradictory advice — here’s a $200 ultra-performance running shoe with a carbon plate as your top rec, but at the very bottom, here’s an $85 model that has more of a lifestyle feel in the completely wrong color. Rinse and repeat. After roughly 20 rounds across five browsers, I arrived at the New Balance 530.

The 530 had also made it to my short list when I’d done the process manually. But while I was faster at narrowing down New Balance models on my own, AI had provided reasoning behind each choice. Things like, this model has extra cushioning for durability or the silhouette in that model would work with multiple outfits. My picks were mostly based on vibes.

Enter phase two: finding a deal. I asked all five browsers to find me the lowest price on a pair of New Balance 530s in all white, white-and-silver, or white-and-pink, in a women’s 8.5 (25cm), that’s in-stock in my zip code. If there was an agentic mode, I asked the AI to put it in my cart.

Cue several more back-and-forths, in which I got differing results. Dia, Comet, Chrome, and Edge found the same local Foot Locker, but selected different colors. Atlas was able to finally put the right pair in my cart, but not without checking in several times to make sure I really wanted to. It also tried to override my preference for pickup and switch to delivery. Once, I watched Atlas spend a minute trying to close a pop-up just to get back to shopping.

I ran the full experiment several times, and each time, I was sure the browsers were finding the best price on a given day. However, I became less and less confident that these were the shoes I really wanted. Especially when Atlas threw in the New Balance 574 Core as an alternative, because they were “one of NB’s most iconic everyday silhouettes” and were a versatile, androgynous shoe. (ChatGPT knows I love unisex styles.)

If I’m judging these browsers on the premise that AI could be better than you at surfing the web, that simply isn’t the case. At no point would I have considered the experience “hands-off.” But more broadly, my whole AI browser experience reinforced that I spend a lot of time doing things for AI so that it can sometimes do things for me. I’m changing the way I think, the way I word questions, and the way I search and digest information. It’s less about how AI fits into my life and more about how I can adapt what I do naturally to accommodate its growing presence.

A good experience with these browsers assumes a lot of things. So does googling, but after 20 years, it requires much less mental effort than the best of what AI browsers currently offer. With AI browsers, you’ve got to be fairly adept at prompting. You’ve got to understand the strengths of chatbots — and be patient enough to work around their weaknesses. Or, at the very least, you have to be willing to learn. This is true for many people. But I’m not confident that anyone who downloads an AI browser will find the learning curve worth it.

AI can sometimes be useful, but it’s always a lot of work. And I still need new shoes. I’ve decided to just visit a New Balance store in person.