Are AI Browsers Any Good? A Day with Perplexity’s Comet and OpenAI’s Atlas

Image by Author

# Introduction

 
Many technology experts will tell you that AI browsers are the future of work. The reality is far more complicated. After spending weeks testing Perplexity’s Comet alongside ChatGPT’s Atlas browser, I discovered something uncomfortable. These tools can save you time on specific research tasks, but they fall apart on anything complex. For data scientists and analysts already overwhelmed by multiple open tabs, this is important. But before you choose Chrome over an AI browser, you need to understand what these tools actually do and, more importantly, what they cannot do.

# Understanding AI Browsers

 
An AI browser sounds like the best browser until you actually use one. At its core, it is a web browser, similar to Chrome or Safari, but with an artificial intelligence (AI) assistant integrated into a sidebar. This assistant can read any webpage you are viewing, answer questions about it, summarize articles, or even attempt to automate tasks — e.g. filling out forms or comparing prices across websites.

Perplexity’s Comet is different from a traditional browser because it understands context across multiple tabs at once. If you have research papers, data dashboards, and competitor reports open simultaneously, Comet’s assistant can read all of them together and synthesize what it finds. The browser was built on Chromium, the same foundation as Google Chrome, so it supports the same extensions you already use.

Image by Author

The key selling point is automation. The Comet Assistant sits in a sidecar panel and can perform multi-tab operations. Instead of manually copying data from five websites into a spreadsheet, you just ask Comet to do it for you.

# Applying Comet in a Data Scientist’s Workflow

 
To understand if AI browsers matter for your workflow, let me walk through what actually happened when a data science professional spent a day with Comet.

// Utilizing Comet for Market Research

If a researcher needed competitive data on machine learning platforms, typically, this involves opening 10 to 15 tabs and manually extracting pricing, features, and performance metrics into a spreadsheet.

With Comet, the researcher created a workspace (Comet calls it a Space) and opened tabs for six competing platforms. Then they asked the assistant:

“Create a comparison table with platform name, pricing per month, free tier features, and latest update date.”

The assistant will read all six websites simultaneously and deliver a properly formatted table in three minutes. This task takes approximately 45 minutes manually.

// Utilizing Comet for Research Synthesis

Next came analyzing three academic papers on anomaly detection algorithms. Let’s say the task is to extract key mathematical concepts, compare methodologies, and identify which approach might work best for their dataset.

With traditional browsing, this means opening each PDF, reading, taking notes, switching between tabs, and manually comparing. They asked Comet:

“Summarize the core mathematical difference between these three papers and rank them by how well they would work with real-time financial data.”

The assistant synthesized across all three open papers and delivered a ranked summary with direct quotes from each source. The assistant even understood the context from their previous questions, so it prioritized financial data applications without being asked twice.

Manual time normally required 2 hours. Actual time: 18 minutes.

// Identifying Comet’s Limitations in Research Synthesis

Here is where Comet breaks down. They needed the assistant to automatically extract data from a JavaScript-heavy dashboard, validate it against a SQL database, and flag mismatches. This is where AI browsers seemingly stop being useful.

Comet could not interact with the dynamic dashboard properly. It tried to click buttons, but the page did not respond as expected. The assistant got confused by the interactive elements and gave up. Our researcher had to do it manually.

This is the honest truth. AI browsers excel at reading and synthesizing static information. They struggle with complex, multi-step interactions in modern web applications.

# Comet Versus ChatGPT Atlas: Which Browser to Choose?

 
If you have been following AI news, you know OpenAI released ChatGPT Atlas, their own AI browser. It performs faster on some tasks but has different limitations.

ChatGPT Atlas is more aggressive with automation. In speed tests, it completed a research brief in 47 seconds compared to Comet’s 58 seconds. For CSV extraction, Atlas was significantly faster, at 36 seconds, compared to Comet’s 52 seconds.
However, Atlas has a critical weakness: it currently only works on macOS. If your team uses Windows or Linux, you cannot use it yet.

More importantly, Atlas is locked into the OpenAI ecosystem. If you want to use it with other tools or workflows, you are out of luck. Comet integrates with Perplexity Pro, allowing it to work alongside your existing AI research workflow.

For data scientists specifically, the choice depends on your primary task. If you are synthesizing information from multiple sources, Comet is better because it maintains stronger context awareness across tabs. If you are automating repetitive form-filling or web scraping tasks, Atlas might save you slightly more time, but only if you use a Mac.

The performance difference is not massive enough to make either browser essential. Both have the same fundamental limitation: they struggle with complex, interactive websites.

# Explaining Why AI Browsers Exhibit Uneven Performance

 
Understanding the underlying technology helps explain why AI browsers exhibit such uneven performance.

When you ask an AI browser to summarize a webpage, it receives the full text content and processes it using a language model. This is essentially what ChatGPT does, but with a single webpage instead of a general prompt. For this task, AI is genuinely strong.

But when a website uses JavaScript to dynamically load content, everything breaks down. The browser can view the rendered page, but the AI cannot necessarily determine which elements are interactive, where the data resides, or how to properly trigger dynamic content. When Comet tried to use the JavaScript dashboard in our afternoon test, it was essentially trying to solve a puzzle it was not designed for.

This is not a flaw specific to Comet or Atlas. This is a fundamental limitation of how modern web applications work. Until AI systems can reliably interpret complex, interactive interfaces in real-time, AI browsers will remain more useful for research and reading than for automation.

# Addressing the Privacy and Security Problem of AI Browsers

 
This is where AI browser companies get defensive. To be useful, AI browser assistants require permission to read your emails, view your calendar, access your contacts, and interact with your accounts. Perplexity and OpenAI need broad access to actually deliver on the automation promises.

That access creates attack opportunities. Cybersecurity researchers have identified a type of attack known as prompt injection attacks as a systemic risk across all AI browsers. Here is how it works: a malicious website hides instructions in its page code that trick the AI into taking actions it should not take. An attacker could potentially manipulate Comet into exposing your emails or making unauthorized purchases.

There is currently no complete defense against prompt injection. It is an emerging problem that the entire industry is grappling with.

For data scientists working with sensitive research data, this is worth considering before granting your browser such extensive access. Many organizations with strict compliance requirements (legal firms, financial institutions, healthcare companies) will not allow these tools on corporate networks because of this risk.

# Real Productivity Gains and Real Limitations

 
Here is what honest testing actually shows.
Tasks where AI browsers deliver measurable time savings include:

• Summarizing long articles or reports (saves 15 to 30 minutes per document)
• Comparing information across multiple static websites (saves 30 to 60 minutes)
• Extracting key information from PDFs (saves 20 to 45 minutes)
• Creating research tables from multiple sources (saves 30 to 60 minutes)

Tasks where AI browsers underperform or fail:

• Working with JavaScript-heavy dashboards or interactive APIs
• Performing multi-step, complex interactions across different sites that require dynamic decision-making
• Handling tasks involving sensitive company APIs or internal networks

One researcher on Reddit who tested Comet intensively reported that it actually doubled their productivity for research synthesis, saving them roughly one hour per day. But this is for someone doing information-intensive work. For general web browsing or working with complex applications, the time savings disappear.

The Perplexity CEO claimed that Comet could boost productivity by 20 percent, potentially adding trillions to GDP. This is venture capital talk, not engineering reality. Real-world usage shows more modest gains — maybe 5 to 10 percent time savings for specific workflows, zero percent savings for others.

# Determining if You Should Switch to an AI Browser

 
The honest answer is, it depends on your actual workflow.
If your work primarily involves reading, analyzing, and synthesizing information from multiple online sources, an AI browser will save you significant time. Data scientists doing literature reviews, competitive analysis, or research synthesis fall into this category.

If your work involves interacting with web applications, filling out forms, or working with JavaScript-heavy tools, it’s best to stick with your current browser. The AI browser will not assist you and may even slow you down.
The cost matters too. Comet requires a Perplexity Pro Max subscription, which costs approximately $20 per month. If you are saving thirty minutes per day on research work, that roughly pays for itself. If you are saving five minutes per week, it does not.

ChatGPT Atlas is currently free during testing, but it will presumably cost something eventually. For macOS users conducting research-intensive work, it may become worthwhile. For everyone else, the equation is less clear.

# Considering Privacy, Security & the Uncomfortable Truth

 
Before adopting any AI browser, you need to understand what you are trading.
AI browsers require sending page content to cloud servers for processing. This means the data you are viewing passes through external systems before being summarized or analyzed. For most research work, this is fine. For sensitive or proprietary information, it is a problem.

Some organizations handle this by using AI browsers only for public research and keeping proprietary work in traditional browsers. This is reasonable, but it also defeats part of the point of having an AI browser.

The prompt injection attack risk is real, but not immediately catastrophic if you use common sense. Do not use AI browser agents to interact with high-stakes websites, e.g., banking portals or sensitive company systems. Use them for research, analysis, and information synthesis where the downside of an error is relatively low.

# Conclusion

 
In conclusion, AI browsers are not going away, but they will not replace traditional browsers anytime soon. The technology will improve. Performance issues with JavaScript applications are likely to improve. New security standards will eventually address prompt injection attacks.

What matters now is matching the tool to the task. For data scientists, the best real-world use case is exactly what we tested: synthesizing research information from multiple sources. This is where AI browsers actually deliver measurable productivity gains without compromising on security or dealing with excessive complexity.

If you spend hours each week reading, comparing, and synthesizing information, test Comet or Atlas for a week. The time savings alone might justify the cost. If your work is primarily application-focused or involves complex interactive tasks, save your money and stick with what you know works.
 
 

Shittu Olumide is a software engineer and technical writer passionate about leveraging cutting-edge technologies to craft compelling narratives, with a keen eye for detail and a knack for simplifying complex concepts. You can also find Shittu on Twitter.