Perplexity AI’s Comet is a Chromium‑based browser released in July 2025 that embeds a large‑language‑model assistant directly into your tabs, promising AI‑native browsing with automation.
Why it matters:
Comet pushes the browser from a passive window to an active collaborator. It can summarize pages, provide cited answers, compare tabs, fill forms, schedule meetings and even add items to carts—all while feeling familiar to Chrome users. Marketing and agency professionals should pay attention because it may streamline research and repetitive tasks.
The big picture:
- Strengths: The browser offers AI search with citations, one‑click page summaries, automation features, @tab context for cross‑page comparisons, voice mode, and privacy controls, all wrapped in a familiar interface .
- Weaknesses: Comet suffers from performance issues, a steep learning curve, limited memory recall, and inconsistent automation that may hallucinate steps. Security researchers have demonstrated “Cometjacking,” a prompt‑injection attack that can exfiltrate data via malicious URLs .
- Outlook: The browser hints at a future where AI agents handle everyday tasks, but it’s still evolving. Teams need clear guardrails around privacy, data access and agent actions before adopting Comet broadly.
What’s next:
Read the full article for a deep dive into how Comet works, its research‑oriented design, real‑world use cases for agency workflows, and a critical look at its privacy and security trade‑offs.
A quiet debut for an ambitious idea
On July 9th 2025, Perplexity AI quietly released Comet – a Chromium‑based web browser with a built‑in large‑language‑model assistant. Comet was pitched as the first AI‑native browser, (even know Dia by The Browser Company released on June 11th, 2025) – a tool that could understand the web page you were on, summarize it and even act on your behalf. Unlike ChatGPT’s Atlas, which sits alongside the page and can click with limited scope, Comet promised a personal assistant that would log into Gmail, fill forms and buy things without you switching tabs.
Marketing and agency professionals might look at Comet as just another shiny tool, but its design hints at a bigger shift. Early AI integrations felt like glorified chatbots grafted onto websites. Comet’s stated goal is to make the browser itself an active collaborator – to shrink the gap between gathering information, analyzing it and acting on it.
What Comet claims to be
Perplexity positions Comet as an AI‑powered web browser built on Chromium, launched in mid‑2025 . The browser is free to use for everyone; a $5 per month Comet Plus add‑on unlocks premium news sources, and a $200 per month Perplexity Max tier offers faster AI responses and priority support . Comet looks and feels like Chrome because it is Chrome under the hood. Users can import their bookmarks, passwords and history with one click and install any extension from the Chrome Web Store, just like in Atlas. The browser works on macOS and Windows, with mobile versions “rolling out,” making it more accessible than OpenAI’s Mac‑only Atlas .
What set Comet apart was the Assistant built into the top‑right corner. You can open an AI sidebar (called the answer box) and ask questions about the current page or any other topic. Comet responds with citations; the assistant will cite sources for its answers . A page summary button instantly produces a short synopsis of the current page . Comet also includes an ad and tracker blocker, accessible via settings, which reduces clutter and increases privacy .
Beyond query answering, Comet is pitched as a do‑the‑thing agent. It advertises the ability to fill forms, book appointments, add items to shopping carts, summarize emails and even reschedule meetings by acting within Gmail and Google Calendar. In an early test, Cybernews asked Comet to find cat food with the most meat content; the assistant navigated the site, evaluated options and added items to a basket . This makes Comet more than a chat overlay. It resembles a general‑purpose operations assistant living in your browser.
Strengths: what Comet gets right today
Familiarity and ease of onboarding
Because Comet is built on Chromium, the interface is instantly recognisable. The browser supports the full Chrome extension ecosystem and imports bookmarks, history, saved passwords and other settings with a single click . There’s almost no learning curve for basic browsing. For teams that already standardize on Chrome, the switch is low friction; there’s less need for re‑training or new workflows.
Integrated AI search with citations
The assistant can be invoked from any page to answer questions, summarize content or provide quick definitions. When asked, it returns answers with citations, allowing the user to verify the information. Traditional AI chatbots often return ungrounded responses; Comet emphasizes transparency by linking back to sources. In the research context, this is invaluable: you can pull the important details while retaining traceability.
Page summarization and context tracking
One of the strongest features is one‑click page summarization. By clicking a three‑line icon in the top bar, Comet produces a concise summary of the current page. This is particularly useful for researchers and analysts scanning long articles or reports. Comet also keeps a limited history of what you’ve asked; you can pick up threads across sessions. The @tab feature allows referencing other open tabs in queries, enabling cross‑document comparisons without copying and pasting – though the implementation is still inconsistent. I do appreciate that you can select multiple tabs, when Atlas currently only allows one in the initial prompt.
Automation tools for routine tasks
Cybernews found that Comet’s automation features can fill forms, book meetings and handle shopping flows automatically. In the cat‑food test, the assistant navigated an online store, analyzed different products, and added items to a basket while showing the user each step. In marketing or operations settings, this could translate to quickly comparing prices, filling out RFP forms or scheduling calls without leaving the browser. I even found success running a daily automation where I have my inbox summarized to me and it highlights the Top 5 emails I need to respond to each morning.
Voice mode and hands‑free operation
Comet supports a voice mode that lets users issue commands without typing, which is also found on Atlas. Voice commands can be useful when multitasking – for instance, reading out an article summary or drafting an email while you’re working on another screen. The same article emphasises that professionals can automate multi‑step tasks like scheduling meetings or summarising emails directly from the browser.
Local and cloud privacy modes
According to independent reviews, Comet offers several privacy modes (local‑only, pseudonymous cloud, and full cloud) that let users choose how much data is processed locally versus in the cloud. This gives control over data exposure. For tasks requiring confidentiality, users can limit the assistant’s access to local processing only, reducing the risk of data leaving the device. Hybrid AI processing aims to balance responsiveness with privacy.
Built‑in ad and tracker blocker
A built‑in ad and tracker blocker can be enabled in settings, providing a cleaner browsing experience and enhancing privacy. While not as comprehensive as dedicated extensions, it offers a baseline level of protection without requiring extra installs. For some corporate environments, integrated protections may simplify compliance.
Where Comet crashes to Earth
Inconsistent automation and @tab reliability
While Comet can perform multi‑step tasks, the execution is uneven. I repeated the Cybernews cat-food test and found that the assistant didn’t check every product on the page; it only analyzed snippets and missed deeper information. The @tab feature – meant to reference specific open tabs – worked inconsistently; sometimes the assistant pulled information correctly, but other times it tagged the wrong tab or required manual rephrasing.
Limited memory and context recall
The promise of remembering what you were working on is not fully realized. In the same review, after adding cat food to a basket, closing the browser and asking Comet to recall the last added item, the assistant could not find any recent basket activity until the user specified the shop. This limitation suggests that Comet’s memory is shallow or not persistent across sessions. For workflows that span multiple days or projects, this reduces its value and you are better off using Atlas where you can continue chats in your preferred ChatGPT client.
Performance and resource consumption
Comet’s agentic features require constant network connectivity and increased CPU usage. Another independent review noted that while pages load quickly, the browser uses more power and network resources than Chrome. This may translate to shorter battery life on laptops and slower performance on weaker machines. In a production environment, these resource demands could become a bottleneck.
Learning curve and reliability for non‑technical users
There are some features that still show performance issues, even a few months after launch and there is a learning curve. At times, I feel like I am using a widget-filled version of a desktop app that I had in my college years in the early 2000s. The automation flows may require specific phrasing; mis‑phrased commands yield no result or cause the assistant to stall. For teams that expect polished tools, this can hamper trust. It is also easier to see the workflow and take over in Atlas currently.
Shallow search and research depth
Comet is built on the Perplexity search engine, which by design surfaces only a handful of results. In practice, this means Comet sometimes surfaces the top three results and misses deeper findings. Researchers may need to leave the assistant and fall back on Google for comprehensive searches, negating some convenience. Not to keep bringing up Atlas, but while Atlas uses a new SERP by OpenAI you can also easily get the results from Google with a single button click. This limitation means Comet remains a supplementary tool rather than a full replacement for search.
Security vulnerabilities and prompt‑injection risks
The greatest weakness comes from Comet’s agentic design. In October 2025, LayerX security researchers published a report demonstrating that a single malicious URL can hijack Comet’s AI assistant, turning it into a data exfiltration tool. The attack, dubbed CometJacking, works by hiding an instruction in the link’s query string. When the user clicks the link, the AI interprets the hidden prompt and begins executing commands. The malicious prompt tells the AI to read from the user’s memory (which may include email, calendar or documents) and encode the data in base64 before sending it to an attacker‑controlled server. Because Comet allows the assistant to access Gmail and Calendar, the attack can steal sensitive data like emails, appointment details and other connected information. Perplexity has basic safeguards to prevent direct data exfiltration, but the researchers showed that encoding data bypasses these checks.
This class of vulnerability is not exclusive to Comet; any agentic browser that reads hidden instructions could be exploited. However, it demonstrates that AI‑native browsers represent a new attack surface. If an attacker can hijack the assistant, they no longer need user credentials – they can instruct the AI to perform actions on their behalf. Security teams will need to treat AI browsers as potential insider threats and implement strict guardrails, monitoring and domain restrictions.
Privacy concerns and data collection
Comet’s ability to act on behalf of the user requires deep integration with accounts such as Gmail and Google Calendar. This raises concerns about how much data the browser collects and how it might be used. The browser offers local‑only modes, but tasks like summarizing emails or scheduling meetings rely on cloud processing. Many potential users worry about giving an AI agent read/write access to personal data – a fear that is exacerbated by the CometJacking vulnerability. Additionally, summarizing content from websites may raise legal questions about republishing or data scraping, especially when publishers rely on page views. This doesn't bode to well for Perplexity as they are currently being sued by Reddit for illegal data scraping.
Real‑world examples: Comet in a marketing or agency workflow
To understand where Comet might fit for professionals, let’s translate its features into common marketing and operations scenarios.
Competitive research and quick briefs
Suppose your team needs to craft a competitor analysis. With Comet, you open each competitor’s homepage in separate tabs. You ask the assistant: “Compare the product positioning and pricing tiers across @tab1 and @tab2; summarize the key differences and cite your sources.” The assistant returns a table with bullet points and links to the specific sections of each site. This reduces manual copy‑paste work and ensures traceability. For a quick internal briefing document, Comet can act as a research aide that speeds up synthesis.
Repetitive form filling and procurement
For tasks like filling RFP forms or requesting quotes from suppliers, Comet can automate the form‑filling process. You can instruct it to fill out company details, contact information and project requirements using stored information. In the Cybernews test, Comet successfully navigated a pet food shop, compared products and added items to the basket. The same approach could help a marketing producer add items to a product comparison spreadsheet or fill out vendor contact forms automatically.
Scheduling and correspondence
Because Comet integrates with Gmail and Google Calendar, an account manager could ask: “Draft an email to our client summarizing yesterday’s meeting; propose moving our next call from Thursday to Friday at 10 AM.” Comet drafts the email and schedules the meeting change, showing you the final text before sending. This reduces friction when juggling multiple client interactions. That said, due to the CometJacking vulnerability, organizations should restrict Comet’s access to sensitive accounts and require human approval before anything is sent.
Content review and summarization
Content teams can use Comet’s summarization feature to quickly condense long articles or research reports. Instead of reading a 50‑page white paper, you can ask the assistant to provide the main argument, supporting evidence and key statistics. It will summarize the text with citations. This is particularly helpful for drafting briefs or blog posts that need to pull from multiple sources.
Campaign ideation and quick inspiration
In an ideation session, you can highlight a competitor’s marketing copy and ask: “What emotional triggers is this campaign using?” Comet analyses the language and identifies themes (e.g. nostalgia, urgency). You can follow up with: “List three other campaigns using similar triggers from the past year.” This helps you find relevant examples quickly without manually combing through search results.
Cross‑department knowledge sharing
Comet’s summary history allows you to revisit previous research sessions (with limitations). A project manager can ask: “Show me the summary we created on the product launch last week and link to the sources we used.” The ability to recall context can improve knowledge transfer within a team. However, as noted earlier, Comet’s memory can be inconsistent. Teams might still need separate documentation to ensure institutional knowledge is retained.
Where Comet might go next
Despite its shortcomings, Comet hints at the future of AI‑enhanced browsing. Here are areas where the product (and the broader category) could evolve:
- Cross‑platform parity – A fully functional mobile version would allow users to move tasks seamlessly between desktop and phone. For marketing teams, mobile parity is critical.
- Granular security controls and audit logs – Businesses will need to define which domains Comet can act on and require approval for high‑impact actions. Browser vendors must implement restrictions on hidden prompt execution and enforce sign‑offs before sending emails or accessing sensitive data.
- Robust memory and project workspaces – To be useful in multi‑week projects, Comet must be able to recall past queries and outputs reliably. Project‑level workspaces with persistent context would increase adoption.
- Domain‑specific workflows and templates – Comet could include playbooks for recurring marketing tasks (e.g., “Competitive snapshot,” “Customer feedback summary,” “Influencer outreach sequence”). Templates would reduce training time and ensure consistent outputs.
- Integration with team collaboration tools – Deeper connections with Slack, Microsoft Teams, Notion or Monday.com could allow Comet to send summaries to channels, create tasks or update databases automatically. This would transform Comet from a personal assistant into a team collaborator.
- Credential management via third‑party vaults – To mitigate security risks, Comet and similar browsers may need to adopt a model where credentials are managed by secure vaults (e.g., 1Password’s agentic autofill). The agent requests credentials; the user approves; but the AI never sees raw secrets. This pattern can protect sensitive logins from being exposed to the model.
Comet’s place in AI Ops
Perplexity’s Comet browser is a bold experiment. By weaving AI assistance directly into the browsing interface, it tries to move research and action into the same workflow. For marketers and agency professionals, the promise is appealing: faster briefs, simplified form filling, smarter email drafts and cross‑tab comparisons without leaving the browser.
But Comet is also an unfinished product. Its automation features are inconsistent; its memory is shallow; its search depth is limited; and its security model has already been shown to be vulnerable to sophisticated prompt‑injection attacks. It remains desktop‑only as of late 2025, and some tasks still require old‑fashioned manual work.
For agencies considering an AI‑first workflow, Comet should be viewed as a prototyping environment, not a production tool. Use it to test how integrated AI might change your processes, but keep your guard up. Define which tasks are safe to automate, require human approval for anything involving money or external communication, and treat the assistant as a junior colleague – helpful, but not infallible.
If Comet improves security, extends context management and offers robust mobile support, it could become a powerful component in a broader AI‑Ops stack. Until then, treat it as a glimpse of the browser’s future rather than the destination itself.
