Who Wins The Browser Wars?

Impressions from my time with Perplexity’s Comet and The Browser Company’s Dia

There is an unholy global tech war rapidly approaching. Applications will fight applications, browsers will fight operating systems, chatbrother against chatbrother. I was given access to two AI browsers—Browser Company’s Dia and Perplexity’s Comet—and saw this future written plainly in the tea leaves.

We can look to the past for clues, but not answers on who will win. In the 2000s, Google Chrome beat every browser in the game, including Microsoft’s Internet Explorer, which had just come off a victory over Marc Andreessen’s Netscape with its superior distribution. Chrome, in contrast, won because it was designed around the company’s insatiable hunger for ads.

More speed? More searches. More users? More data for ad targeting. As Charlie Munger famously said, “Show me the incentive and I’ll show you the outcome.” The outcome is that, today, Chrome has completely outstripped its past competitors.

But ads don’t work as well in this new era. Instead, the goal is the total ownership of your mind. It’s not just Dia and Comet up for this fight. Reuters also reported this week that OpenAI is looking to release a browser in the upcoming weeks.

As a disclosure, both companies gave me early access but in no way tried to influence my results. They both just wanted to know what I thought. (I have 5 Dia invites to share. If you want to try it out, respond to this email and I’ll give them to the first five paying subscribers to The Leverage.)

Vectors of competition

Chrome’s big innovation was speed. Up until LLMs became popular, being faster meant loading web pages faster than your peers. When AI truly integrates with browsers, the speed browsers are competing on is speed to complete a task, not loading speed. I think of task completion as competing along two vectors:

1. Do stuff for (Application Agents): All of us access individual applications—websites—via a web browser. Most popular websites are attempting to be the system of record for their chosen market, the central hub holding the most crucial data and workflows. I’ll go to LinkedIn to look for a new job (and kill time), X to network (and kill braincells), etc. The AI browsers I’ve tested let me outsource specific tasks within these applications. I may tell it to “write a viral tweet” and then the agent will do stuff for me. Comet has an integration with Gmail and Google Calendar so it can answer emails or change my calendar too (it is a somewhat terrifying proposition to give my inbox over to a robot and one I can’t quite stomach yet).

   The models can also be given a general task, like: “I am going on a date night with my wife. We want to see the new Wes Anderson movie and get a burger. Please plan this out, buy the tickets, and make a reservation.” The agent will open up several tabs of its own accord. Then, it will act as you across individual applications to accomplish the task. “Do stuff for” is the specialization of Perplexity’s browser Comet.

2. Do stuff with (Multi-Tab Agents): Outside of application specific labor, there is something I would deem as “general internet usage” where you are reading, researching, or searching for information. This is stuff where you type in a question into the search engine or chatbot of your choice.

   In my life as the founder of The Leverage, to prepare for a piece like this, I use my old, dumb browser to painstakingly source and read dozens of essays about the history of the internet browser, and maybe watch a video or two about the time period. I (should) take notes on what I learn, and then compile that into some kind of essay for all of you. My sense is that this process serves me beyond mere data collection. While I may grab a specific number or two—such as for the market share chart above—much of this reading is meant to train my subconscious mind on the task at hand. “Do stuff with” agents make the process of sourcing, accessing, and fact-checking against this material easier. They combine and remix the data contained in multiple tabs.

   Right now, I’m working on an essay about how AI alters our state of being, with lots of references to the German philosopher Martin Heidegger. Using Dia, I can simultaneously chat with my notes, the outline I’ve prepared, and the various essays I’m reading to ask questions like: “Does my thesis align with Heidegger’s thesis around relational truth?” To be fair, you could probably do this by adding all of this data into a prompt on ChatGPT, but it is faster and easier to do it via an AI browser. Remember—speed, speed, speed is what counts. This AI browser use case is what Dia is currently focusing on.

In both Dia and Comet, you primarily interact with the AI through a chat menu on the right hand side of your screen that you can toggle on. This, however, is where you start to get hints of the problems that come with this product category. Every application that I use today also has an AI chat window. When you combine these website’s chat windows with Dia’s vertical tabs, you get something hilariously ugly. Below, I’ve got writing app Lex open inside my Dia browser. It is worth zooming in on this so you can see how funny it looks.

Source: Author’s screenshot

The only place that work is actually happening is in the second panel. The third panel is Lex’s AI chatbot that is deeply integrated into my writing workflow. Dia’s AI chat is on the farthest right and has context across all of my open tabs. Eek. AI browsers have to either offer a deeper AI integration into my existing application specific workflow (unlikely) or have context from my previous browsing or chat history that makes its weaker integration more powerful. Keep in mind that both the browser and the application are using the same generation of models from OpenAI and Anthropic!

Ugly UX aside, there’s a larger problem with any product—browser or something else—offering AI agents. They, uh, don’t really work. Like at all. With Comet, I was only successful with maybe 50% of my more complicated tasks. The only place that AI agents have found product market fit, is with AI coding tools. Those environments are successful because, as I’ve argued before, they have:

1. Testable results: AI agents for coding have definitive benchmarks. Does the code run, is it error-free, and does it meet the goals of your specified task? These benchmarks provide immediate and objective feedback for agent performance.

2. Very clear pass/fail scenario: Software coding has discrete criteria. Code either compiles or it doesn't. Tests pass or fail. Thus, success and failure states are transparent and easy to measure.

3. Large data set for training (public code) and fine-tuning (existing code base): Coding agents tap into an enormous publicly available corpus (GitHub) alongside a company’s proprietary internal codebases, making it straightforward to train and continually refine AI accuracy.

4. Clear ROI: The productivity gains from automating coding tasks are immediately calculable in terms of reduced developer hours, faster deployment cycles, and direct cost savings, making ROI straightforward to communicate and justify to buyers.

5. Layered safety nets: Software development allows easy implementation of layered oversight, such as human reviewers, automated tests, continuous integration pipelines, and rollback capabilities. This makes agent errors less of a big deal because they are caught before they get too far.

An AI agent successfully navigating even a single individual application would require an absolutely cracked engineering team, focused 100% of their time on making it functional. They’d need to reverse engineer and map websites, update as applications changed, and understand a user’s intentions extremely well. It is enormously challenging for either Perplexity or The Browser Company to deliver an AI agent that works for the entire internet of millions of websites. That is far, far outside the reach of any model today.

That does not mean that these products are bad—far from it. When agents are successful, I make a little cooing expression, and go ooooooooh to my computer. It is so impressive. What’s unclear to me is what ratio of magic to error the average user is willing to accept. The Perplexity team told me that their waitlist had 500K+ people on it as of July 10th, so it’ll likely be the same balance as me for some of these users.

So too, with Dia. When Dia combines multiple tabs and gives me insights, my body is filled with a genuine wonder. However, the system is still limited by the weaknesses of the LLM. Heidegger is notoriously, ridiculously dense. He has a nasty habit of inventing new words or inventing new definitions of old words, so you have to pay incredibly close attention to each thing he says. I have often found that Dia (and ChatGPT) are making subtle errors that would be unnoticeable to anyone who hadn’t studied this stuff. These errors stem from the underlying intelligence and hallucination rate of the models powering the experience, and there is relatively little either The Browser Company or Perplexity can do to correct that outside of training their own models. This, of course, is unreasonably expensive at the moment, so that is unlikely to happen.

For both companies, the bet, which is reasonable, is that the models will catch up to the abilities that the browsers are designed for. The reason why all these companies are releasing browsers now takes us back to our original question of incentives.

Implications and incentives

Both The Browser Company and Perplexity appear to monetize primarily via the incineration of venture capital. The Information reported that Perplexity had -$69M in net income last year while The Browser Company has never, as far as I can tell, had a paying customer.

Let’s assume that the same monetization dynamics that existed for Chrome hold true today. Google originally wanted a browser that felt fast so people would search more. They also wanted the data a browser could provide to target their ads better.

AI browsers have slightly different dynamics. As we’ve previously discussed, AI companies' record-breaking speed of revenue growth demonstrates that their customers—both consumers and enterprise buyers—will pay for labor automation. I see no reason why browsers would break that trend. To get access to Comet today, you can join Perplexity’s Pro Tier which costs $200 a month. Just on the basis of its 500K+ waitlist, at least some people are willing to pay that price. The case here seems fairly straightforward. Save me lots of time and effort and I’ll pay you for that.

Advertising is slightly more complicated. In the previous generation of the internet, Google was able to offer services like Chrome for free because of its positive ROI on ads. A user would click around, see ads, and Chrome would make money off this process. Pretty simple. In an AI browser world, that process of clicking around will be executed and decided on by agents rather than human beings. What do ads for agents look like? Do they even exist? No one knows the answer. Will you have ads inserted into the middle of your workflow? Will you have to watch an ad before your AI agent does a task? When you add in the fact that AI features cost GPU time, it feels likely that consumer subscriptions will be the monetization model for the next few years—at least until we figure out what ads might look like in this new world.

In which case, it makes sense to rush these products. Browsers are gathering the new oil of data. This data will help them create a competitive moat of user preferences expressed in three ways:

1. User history: What are the most common sites that a user frequents? How long do they spend on these sites? At what hours do they visit?

2. User profiles: What kind of taste does this user have for what they watch and read? What is their job and preferred mode of communication?

3. Common prompts: Dia calls these “skills.” They refer to prompts that you want to run frequently enough to have a hotkey for. These hotkeys can tell you a lot about a person and what future features to build.

Together, these three data streams make it exponentially easier to get the next generation of models to execute on agentic tasks. These companies are incentivized to get their browsers out there so they can build up the data set that frees them from having to build models that cost billions of dollars.

Be prepared. If (when) AI agents can fully execute day-to-day knowledge work, much of the software that we have today will be rendered irrelevant. All that will matter is the unique data they have contained within them. The products we have discussed today bet on neatly sidestepping that fight by being the portal by which we, and our new AI agent servants, access these suddenly irrelevant applications. If we access the world through a browser, so can our AI agent, and whomever controls that agent controls the majority of the value.