Executive Summary
Comet browser features explained in one sentence: Perplexity has taken a Chromium browser and exposed selected browser context to an AI assistant that can read, compare, summarise, navigate, and act. The striking part is that the same context that makes Comet genuinely useful also creates its hardest privacy and security questions. I therefore do not treat Comet as “Chrome with a chatbot”. I treat it as three systems sharing one window: a conventional browser, a context-aware research layer, and an agent that can control web interfaces when I give it permission.
That distinction matters in 2026 because Comet now spans desktop and mobile, supports most Chrome extensions on desktop, adds page and selected-text assistance, offers reusable prompt shortcuts and voice navigation, and can move from reading a page to executing multi-step work. Perplexity’s current product page lists Mac, Windows, iOS, and Android availability, while its Help Centre documents enterprise policy controls, local device storage, prompt-level permissions, and a subscription model that separates browser access from higher-volume agent usage.
This guide focuses on what a serious user needs to know before making Comet a default browser: which features are actually documented, where mobile and desktop differ, what data is sent, how long context may be retained, where agent quotas become a planning constraint, why prompt injection is not solved by better prompting alone, and which workflows create enough time savings to justify delegation. Where Perplexity’s own documentation conflicts or does not publish an exact limit, I call that out instead of filling the gap with an estimate.
The Browser Layer Comet Actually Adds
Comet begins with a familiar technical choice: Chromium. That gives it the rendering engine, tab model, bookmark behaviour, password and autofill patterns, and extension ecosystem that make migration from Chrome relatively low-friction. Perplexity’s current Getting Started with Comet page confirms standard features such as bookmarks, page translation, and support for most Chrome extensions. The product difference is the AI layer placed directly against browser state, not a separate chat page that requires manual copy and paste.
The assistant can work with visible page text, headlines, and metadata, answer questions about the current page, and use a named tab as context. Comet also exposes higher-level controls such as natural-language navigation, tab management, reusable shortcuts, and approved browser actions. This is the architectural shift described in our agentic workflow overview: the browser is no longer only a renderer and navigation shell. It becomes a tool environment in which an AI model can observe browser state and, under defined permissions, alter that state.
| Feature Layer | Verified Capability | Important Boundary |
| Browser core | Chromium engine, tabs, bookmarks, translation, history, profiles, password and autofill functions | Chrome-specific integrations are not guaranteed to behave identically |
| AI search | Perplexity-powered answer experience and cited synthesis on supported surfaces | A citation still needs claim-level verification for consequential decisions |
| Page context | Summaries, extraction, questions about visible content, selected-text assistance | Context is sent only when the request needs it, but qualifying context may be retained |
| Cross-tab work | Reference named tabs, compare pages, extract actions, summarise several tabs | More tabs increase irrelevant context and prompt-injection exposure |
| Browser control | Navigate, click, type, fill forms, organise tabs, and run multi-step tasks | Sensitive or irreversible steps should keep a human approval point |
| Enterprise control | MDM deployment, 500+ Chromium policies, domain-level Assistant permissions | Enterprise controls do not remove the need for least-privilege task design |
The most useful mental model is a permission ladder. A page summary asks Comet to read one surface. A multi-tab comparison widens that surface. A history search adds personal context. A connected mailbox adds authenticated information. Browser control adds the power to act. Those are not five variations of the same feature. They are five different trust levels. The feature set becomes easier to manage once each task is placed on that ladder before the assistant starts.
Assistant Context: Pages, Tabs, Text, and Memory
The Assistant panel is the centre of Comet’s contextual browsing model. Perplexity documents it as a side panel that can answer questions, summarise a page, extract dates or action items, draft text from page content, and pull in another tab when that tab is explicitly referenced. This is where Comet feels materially different from a normal browser extension because the interaction is designed around browser state rather than a single pasted document.
Start With the Smallest Useful Context
For reliable work, scope should expand only when necessary. Selected text is narrower than a full page. The current page is narrower than a named set of tabs. A named set of tabs is narrower than history or connected services. When I want a clean answer, I begin with extraction: “List every price shown on this page and preserve the plan names.” I only ask for interpretation after I have verified those extracted values. The browse the web with Comet guide develops this read, compare, verify, then act pattern for deeper browsing sessions.
Inline Assistant makes that narrow-context workflow faster. Selecting text on a page or in an input field opens quick actions such as summarise, fact-check, define, translate, improve writing, or ask a custom question. Perplexity notes that Inline Assistant is unavailable in Incognito and can be blocked on some sites. That limitation is useful because it reminds users that AI assistance is not an invisible system-wide capability. It operates where browser and site permissions allow it.
Memory Is Not the Same as Browser History
Comet’s memory model is easy to misunderstand. The browser itself does not maintain a separate AI memory store. Perplexity says Comet uses the same account memory system as Perplexity, including saved memories and the Search Library, and that agent actions such as browsing a site or completing a task are not stored as memories. Browser history, cookies, cached files, and autofill data are a different layer that is primarily stored on the device unless the user enables relevant sync features. The practical result is that “remember what I prefer” and “find the page I visited last week” can invoke different data paths.
Commands, Shortcuts, and Voice as an Interface
Comet’s most important interface change is that navigation can be expressed as intent rather than a sequence of clicks. A user can ask the browser to find a page, organise tabs, compare sources, or continue a task while browsing elsewhere. The Assistant panel supports parallel errands, while query shortcuts turn frequently repeated prompts into reusable commands. Perplexity describes shortcuts as mini AI agents that can point at files, tabs, and @-mentioned sources, then combine several steps under one trigger.
A good shortcut is bounded and inspectable. “/weekly-pricing” can specify the vendors, official sources only, the fields to extract, and a comparison format. A poor shortcut is “research competitors”, which leaves source selection, freshness, scope, and stopping conditions undefined. Reusability magnifies whatever is in the prompt. If the prompt is precise, the shortcut saves time. If the prompt is vague, the shortcut industrialises ambiguity.
Voice Adds Speed, Not Automatic Precision
Perplexity’s February 2026 Comet update says Voice Mode is powered by GPT Realtime 1.5 and can use screen awareness, navigate the web, take actions by voice, and preserve conversational context as the user switches tabs. The documented desktop shortcut is Option+Shift+V on Mac and Alt+Shift+V on Windows. Voice is particularly useful when the user is reviewing a dashboard, cooking from a recipe, preparing meeting notes, or walking through a long page without wanting to type.
The risk is referential ambiguity. Spoken prompts naturally contain words such as “this”, “that number”, and “the other tab”. For low-stakes reading, the convenience is worth it. For research or action, name the page, company, date, and metric. Voice should reduce input friction without reducing the specificity of the instruction. I would rather say one extra sentence than let an assistant choose which authenticated tab “send this” refers to.
Automation: What Comet Can Do for You
Comet becomes an agentic browser when the assistant moves from describing the web to changing it. Official examples include navigating sites, clicking, typing, filling forms, managing tabs, working through email and calendars, comparing products, and carrying out multi-step errands. Enterprise documentation separates Comet Assistant, the reading and research companion, from Comet Agent, the executor that can complete workflows such as finding flights against a calendar, scheduling meetings, filling forms, and managing subscriptions.
The right implementation rule is to separate reversible preparation from consequential execution. Asking Comet to draft an email, prepare a booking form, identify three compatible products, or place a proposed meeting on a checklist is reversible. Sending the email, confirming the booking, purchasing the product, deleting an account, or accepting legal terms is not. Our guide to automate online tasks with Comet goes deeper on setting approval points so an agent can remove repetitive work without receiving unnecessary authority.
| Workflow | Good Delegation Boundary | Human Checkpoint | Typical Failure Mode |
| Research comparison | Collect and structure evidence from named sources | Verify decisive claims and dates | Source mixing or stale documentation |
| Summarise threads and draft a reply | Review recipients, claims, tone, attachments | Wrong thread context or premature sending | |
| Calendar | Read availability and propose slots | Approve attendees and final invitation | Timezone, recurrence, or participant mismatch |
| Shopping | Compare price, availability, delivery, and returns | Approve product and checkout | Unavailable item, misleading merchant page, changed price |
| Forms | Pre-fill known, non-sensitive fields | Review sensitive data and submission | Hidden validation, ambiguous field, anti-bot challenge |
| Tab management | Group, close, or reopen clearly defined tabs | Confirm destructive bulk actions | Closing a needed unsaved form or session |
Large-scale field evidence supports the idea that people value these practical categories. A 2025 study by Jeremy Yang and colleagues, based on hundreds of millions of anonymised Comet interactions, found that Productivity & Workflow plus Learning & Research represented 57% of agentic queries. Personal use accounted for 55%, professional use 30%, and educational use 16%. Those figures do not prove that every delegated task saves time, but they show where real users concentrate agent behaviour rather than where product demos merely look impressive.
Comet Browser Features Explained for Research Work
Research is where Comet’s browser context produces the clearest advantage over a standalone chatbot. Instead of moving material into a separate conversation, a researcher can open primary documents, ask page-specific questions, compare selected tabs, extract dates or claims, and then use Perplexity search to broaden the evidence set. The key is to make the browser preserve disagreement rather than compress every source into one confident paragraph.
Use an Evidence Contract
A strong research prompt defines five things: the decision being informed, the date boundary, the allowed source classes, the required counter-evidence, and the output structure. For example: “Compare the current enterprise prices of these services using only official vendor pages updated in 2026. Show monthly and annual price, published caps, missing information, and the exact source used for each row.” This approach turns Comet from an answer generator into an evidence organiser. The step-by-step research a topic with Perplexity workflow adds source role labelling and verification passes for longer investigations.
One useful research pattern is a contradiction ledger with columns for claim, supporting source, conflicting source, and unresolved question. Another is a freshness split: ask separately for the newest official documentation and the newest independent reporting. Product documentation is best for current feature state and plan terms. Independent reporting and security research are better at surfacing trade-offs, historical failures, or patches that changed the story. Keeping those streams separate prevents a 2025 vulnerability from being written as if it were definitely unpatched in 2026, while also preventing a new vendor page from erasing the significance of the original disclosure.
A July 2026 preprint by Yan Xia and colleagues audited 13,777 articles and 41,331 summaries across Chrome with Gemini, Edge with Copilot, and Perplexity Comet. The authors found the browser summaries broadly accurate overall, while also reporting systematic changes in ideological tone, negativity, anger, fear, clarity, and personal voice. That is a useful reminder that summarisation can be factually sound and still alter how a reader experiences a source.
“AI-powered browsers are a new class of editorial intermediaries that systematically reshape news.”
Yan Xia and co-authors, 2026 preprint on browser-based AI news summarisation.
Compatibility: Extensions, Profiles, Sync, and Mobile
Chromium compatibility makes Comet easier to adopt than an entirely new browser engine would be. Perplexity says most Chrome Web Store extensions work on desktop, and users can import bookmarks, passwords, cookies, history, and extensions from supported browsers. Profiles can separate work and personal browsing spaces, including their own history, bookmarks, passwords, extensions, theme, and cookies. That separation is valuable for AI browsing because it reduces the chance that an assistant task performed in one context sees an unrelated authenticated session in another.
Extension support should still be treated as a permission problem, not only a compatibility checkbox. An extension that can read and change data on many sites already has substantial power. When the same browser also hosts an assistant with page context and browser-control capabilities, the combined trust surface grows. Our browser extension permissions guide explains how to inspect host permissions, developer reputation, and whether an extension duplicates a feature Comet already provides.
Desktop Requirements Are Documented Inconsistently
Perplexity’s official desktop documentation currently contains a technical inconsistency worth flagging. The general Installing Comet page lists Windows 10 or later, macOS 11 or later, 4 GB RAM minimum with 8 GB recommended, and 500 MB of available disk space. A separate Operating System Requirements page lists Windows 10 and 11, macOS 13 or later, the same 4 GB minimum, but at least 2 GB free with 5 to 6 GB recommended. Until those pages are harmonised, organisations should plan against the stricter requirements rather than the looser installer summary.
Mobile documentation has also moved quickly. The current Comet product page lists Mac, Windows, iOS, and Android. Older Help Centre pages can still surface claims that no mobile app is available, while newer Android documentation and Perplexity’s March 2026 iOS launch material confirm the mobile rollout. This is a practical example of why “latest official page” matters more than a search snippet. Mobile feature parity is not complete: desktop-style extensions are not necessarily present, and the interface favours voice, search, page summary, and task actions.
“Google does a much better job than anyone else in the world, including Perplexity.”
Aravind Srinivas, Co-founder and CEO of Perplexity, explaining the Google default for many iOS navigation and local searches in March 2026.
Privacy: What Stays Local and What Gets Sent
Comet privacy is easiest to understand by separating ordinary browser storage from Assistant requests. Perplexity’s March 2026 privacy guidance says the Assistant does not access or upload browsing history, the full list of open tabs, cookies, site data, passwords, autofill data, local files, or text typed into websites by default. Those categories remain local unless a user requests a feature that needs relevant context or enables a sync path. A page summary, selected-text extraction, @tab request, history-based task, or email/calendar workflow can therefore send the minimum context needed for that request.
The retention detail is more important than the marketing shorthand “local by default”. If an Assistant request needs page content, history items, or open-tab context, Perplexity says that context can be stored securely for up to 30 days to support Library and query-history functions. Threads using that context are treated as temporary threads, where users can delete them or change the expiry behaviour. The browser can therefore be local for ordinary browsing data while still creating temporary server-side context for an AI task.
| Permission Layer | Example | Recommended Default | Why It Matters |
| Selected text or page read | Summarise a passage or current page | Allow on ordinary public pages | Smallest contextual surface and easiest evidence check |
| Named tabs | Compare two product pages or reports | Limit to only task-relevant tabs | Each extra tab adds data and possible hostile instructions |
| History context | Find a page visited last week | Use on demand | May expose personal browsing patterns when explicitly requested |
| Connected services | Read email or calendar context | Connect one service at a time | Authenticated information raises the impact of a mistake |
| Browser control | Click, type, navigate, fill forms | Allow once for bounded tasks | Changes browser state and can cross site boundaries |
| Consequential action | Send, submit, buy, delete, reschedule | Require final human review | Errors may be costly or difficult to reverse |
Comet also documents built-in ad and script blocking, fingerprinting protection, safe-browsing warnings, site permissions, Incognito behaviour, and local password decryption after operating-system verification. Third-party cookies are allowed by default for compatibility, which is a meaningful trade-off rather than a hidden surprise. Users who prefer stricter tracking protection can change that setting, but embedded logins, payment widgets, and other site functions can break. The broader complete Perplexity AI guide covers account-level controls and search history outside the browser itself.
Security: Prompt Injection Is the Hard Problem
The most serious limitation of agentic browsing is not that an AI might summarise a paragraph badly. It is that the model consumes untrusted web content while also possessing tools that can read private context or act in authenticated sessions. An attacker can place instructions inside a webpage, document, comment, image, or other content and try to make the agent treat those instructions as if they came from the user. This is indirect prompt injection, and it changes the browser threat model because content and control can enter the same model context.
Trail of Bits disclosed in February 2026 that Perplexity had hired the firm before Comet’s launch to adversarially test the browser. Its audit demonstrated four prompt-injection techniques capable, in proof-of-concept conditions, of exfiltrating Gmail content through the Assistant. The important current fact is not to imply those pre-launch demonstrations remain unpatched. Perplexity says it addressed the identified gaps and used the audit to shape ongoing testing. The lesson is architectural: high-authority agents need defence in depth, explicit trust boundaries, and controls that do not depend only on a model deciding that suspicious text “looks unsafe”.
“Trail of Bits’ systematic approach helped us identify and close these gaps before launch.”
Kyle Polley, Security Lead at Perplexity, quoted by Trail of Bits, February 2026.
Perplexity’s BrowseSafe research adds a measurable defence layer. Its specialised classifier reached about 0.91 F1 on the published benchmark, compared with roughly 0.85 for frontier GPT-5 and Sonnet 4.5 configurations and about 0.35 for PromptGuard-2. The research also acknowledges a production constraint: detection must be fast enough not to bottleneck every agent observation. More importantly, the paper advocates layered protections rather than a single detector, including architectural isolation and model-based safeguards.
“The fundamental vulnerability is identical: the collapse of the instruction/data boundary inside a shared context window.”
Ali Shahin Shamsabadi, Hamed Haddadi, and Artem Chaikin, Brave security researchers, June 2026.
For users, the practical controls are simple even if the underlying problem is not: keep the Assistant blocked or read-only on sensitive domains, use separate profiles, prefer allow-once permissions, close unrelated authenticated tabs before broad multi-tab tasks, and review the final action. Enterprise teams gain stronger policy tools, including Browser Control, Read Only, and No Access by domain, plus more than 500 Chromium-based policies delivered through MDM. These controls reduce exposure, but no responsible evaluation should describe an agentic browser as risk-free.
Pricing and Limits: The Cost Is Mostly Quota Ambiguity
Comet can be downloaded and used for ordinary browsing without buying a premium Perplexity subscription, but the useful distinction is between basic browser access and sustained Assistant or agent usage. Perplexity’s July 22, 2026 plan comparison lists Standard, Pro, Education Pro, Max, Enterprise Pro, and Enterprise Max. Pro is currently $20 per month or $200 per year, Max $200 per month or $2,000 per year, Enterprise Pro $40 per month or $400 per year per active seat, and Enterprise Max $325 per month or $3,250 per year per seat. Education Pro is listed at $10 per month with verification.
| Plan | Current Price | Browser Agent Position | Published Limits or Caveats |
| Standard | $0 | Browser Agent queries listed as unavailable in plan comparison | 3 Pro Searches/day; 1 Research query/month; basic browsing and limited AI still available |
| Pro | $20/month or $200/year | Monthly limits described as average use | No fixed public Browser Agent query count in the comparison table |
| Education Pro | $10/month with verification | Monthly limits described as average use | Education eligibility required; exact agent count not publicly fixed |
| Max | $200/month or $2,000/year | Monthly limits described as advanced use | 10,000 monthly Computer credits; exact Browser Agent count not publicly fixed |
| Enterprise Pro | $40/month or $400/year per seat | Extended usage | 80 Browser Agent queries/month; 400 Pro Searches/week; 50 Research queries/month |
| Enterprise Max | $325/month or $3,250/year per seat | Highest enterprise usage | 800 Browser Agent queries/month; 4,000 Pro Searches/week; 500 Research queries/month |
The hidden cost is not an undisclosed surcharge. It is planning uncertainty. Enterprise plans publish numeric Browser Agent caps, while consumer Pro and Max use phrases such as “average use” and “advanced use”. That may be adequate for an individual whose workload can flex, but it is weak input for a team trying to calculate a cost per automated task. Computer credits are a separate meter again: current documentation lists no recurring monthly Computer credit allocation for consumer Pro, 10,000 monthly credits for Max, 500 per month for Enterprise Pro, and 15,000 for Enterprise Max, with promotional bonuses and purchased credits subject to their own rules.
API access is also separate. A Perplexity web subscription does not include programme-level API usage, and there is no public Comet browser-control API positioned as a substitute for the user-facing Agent. If a company needs deterministic, developer-managed automation, it should price the Perplexity API or another automation stack separately rather than assuming a Comet subscription is an API licence. For teams choosing between browser-led automation and conventional browsing, our Comet and Chrome comparison focuses on the operational trade-off between AI context and mature browser predictability.
Performance Bottlenecks and Failure Modes
Comet can feel fast because it removes manual context switching, but “fewer clicks” is not the same as “lower task latency”. A single-page summary may complete quickly. A multi-tab task that retrieves sources, reasons over them, invokes a connector, encounters a login wall, and then manipulates a form can be slower than a competent human performing one obvious action. The useful unit of performance is therefore end-to-end completion time including corrections, not model response time.
The common bottlenecks are predictable: ambiguous instructions, large or noisy tab sets, slow sites, anti-bot challenges, expired authentication, unsupported page elements, extension conflicts, changed selectors, and sites that require a human judgement the agent cannot infer. Context itself can become a bottleneck. Perplexity’s 2026 research on query-aware context compression describes “context rot”, where irrelevant material consumes model capacity and can make the answer less focused. In practical browser work, opening ten vaguely related tabs can therefore make the AI worse at the exact comparison the user cares about.
Recovery Is a Feature Requirement
A good automation plan includes a failure state before execution. Tell Comet what to do if a page blocks automation, a field is ambiguous, or a required value is missing. “Stop and ask me” is better than letting an agent improvise a legal name, payment choice, shipping address, or calendar assumption. For research, specify that missing evidence must be labelled unavailable rather than inferred. For shopping, prohibit substitution unless explicitly approved. For email, require a draft only when the thread contains contradictory instructions.
Troubleshooting should start with the browser layer before blaming the model. Confirm Comet is current, retry the task in a fresh tab, disable or isolate suspect extensions, verify site permissions, check whether the service is reachable in a conventional browser, and reduce the task to one action. The Comet troubleshooting guide covers update, profile, extension, and service-level checks. If the reduced task works, rebuild the workflow one step at a time so the failure point becomes visible.
Who Should Use Comet and Who Should Stay With a Conventional Browser
Comet is a strong fit for people whose browsing consists of repeated reading, comparison, synthesis, and light execution. Analysts can keep source material open while extracting structured facts. Recruiters can summarise a role page and prepare outreach. Students can compare explanations and ask follow-up questions without moving between a browser and a separate chatbot. Operations teams can prepare repetitive forms or scheduling work with approval gates. The value rises when browser context is genuinely useful and the user already has a verification habit.
A conventional browser remains the better default for work where the assistant adds little but the permission surface adds a lot. Banking, healthcare portals, password-management consoles, administrator dashboards, regulated legal workflows, and high-value purchases deserve a conservative posture. So do environments that depend on specialised Chrome extensions or corporate controls not yet validated against Comet. The existence of enterprise MDM and policy support is promising, but migration should still begin with a controlled pilot rather than a company-wide default-browser switch.
The decision is also task-specific rather than ideological. On mobile, Perplexity CEO Aravind Srinivas publicly acknowledged that Google remains stronger for many navigation and local searches, which is why Google is used as the default provider for those iOS flows. That is a useful example of the balanced view an AI-browser evaluation needs. Perplexity does not need to be best at every retrieval task for Comet to be valuable. Its advantage is the orchestration layer around browsing, context, reasoning, and action.
My practical threshold is simple: use Comet when it removes at least two meaningful manual stages and the result can be checked before harm occurs. If the task is one click, one sensitive decision, or one action that cannot be easily reversed, the AI layer may be unnecessary. The best browser is the one that reduces cognitive work without hiding where evidence ends and automation begins.
Our Content Testing Methodology
This article was built from a source-first verification pass completed on 11 August 2026. I used Perplexity’s current Comet product page and Help Centre for feature state, privacy behaviour, enterprise controls, plans, quotas, and system requirements. Where two official pages disagreed, such as desktop operating-system and disk-space requirements, I reported the conflict and used the stricter specification for planning guidance rather than inventing a reconciliation. I also distinguished current product availability from stale Help Centre pages that still describe an earlier mobile state.
Security claims were cross-checked against Trail of Bits’ February 2026 Comet audit, Perplexity’s BrowseSafe research, and Brave’s June 2026 analysis of indirect prompt injection. Adoption and research-use statistics came from the large-scale Comet usage study by Yang and colleagues, while the news-summary discussion uses the July 2026 preprint by Xia and colleagues. No live interactive Comet automation session was performed for this draft, so I do not present unperformed hands-on tests as evidence. Workflow recommendations are derived from documented capabilities, published audits, and reproducible least-privilege practice.
For internal linking, the live XML sitemap did not return parseable XML through the available browsing layer. To avoid fabricating sitemap entries, I used the brief’s fallback rule and verified eight live Perplexity AI Magazine pages directly, then placed each internal URL once in a different body section with contextual anchor text. No internal links appear in the Introduction, Executive Summary, FAQs, or Conclusion.
Publication disclosure required after human desk review: “This article was researched and drafted with AI assistance and reviewed by the Sami Ullah Khan editorial desk at Perplexity AI Magazine. All data, citations, pricing figures, and named quotes have been independently verified against primary sources before publication.” The quoted disclosure should only be treated as complete once the named editorial desk has performed that final human review.
Post-publication technical QA should include a normal browser back-button test and an inspection for hidden text or accidental display:none content in the WordPress rendering. Those checks are sound publishing controls in their own right. This research pass did not independently verify the specific enforcement dates stated in the production brief, so no date-specific Google enforcement claim is presented here as fact.
Conclusion
Comet’s defining feature is not a single button. It is the decision to let an AI assistant work against browser context and, when authorised, browser actions. That produces real advantages for research, cross-tab synthesis, repetitive workflows, voice-driven navigation, and tasks that would otherwise require constant movement between a browser, search engine, and chatbot. It also creates a permission and security model that conventional browsing does not require.
The current product is more mature than its early invite-only reputation suggests: it is available across major desktop and mobile platforms, supports most Chrome extensions on desktop, includes enterprise deployment controls, and has documented privacy boundaries. Yet the unresolved issues are meaningful. Consumer agent quotas remain less transparent than enterprise quotas. Official technical documentation is not perfectly synchronised. Indirect prompt injection remains a structural challenge for the entire agentic-browser category. Mobile feature parity is still evolving.
For 2026, the sensible position is neither “replace every browser with Comet” nor “AI browsers are unusable”. Comet is most valuable when context reduces real work and a human remains responsible for evidence, permissions, and irreversible actions. The open question is how quickly security isolation, quota transparency, and cross-platform consistency can catch up with the speed at which browser agents are gaining authority.
Frequently Asked Questions
Is Comet Browser Free to Use?
Yes. Comet can be downloaded for ordinary browsing and basic AI use without a paid Perplexity plan. Higher-volume Assistant, advanced-model, Research, and agent capabilities depend on the user’s Perplexity subscription and published usage limits.
What Are the Main Comet Browser Features?
The core features are Chromium browsing, Perplexity-powered AI search, a contextual Assistant panel, page and selected-text summaries, cross-tab questions, natural-language browser commands, reusable shortcuts, voice navigation, browser automation, Chrome extension support on desktop, privacy controls, and enterprise policies.
Is Comet Based on Chromium?
Yes. Perplexity documents Comet as Chromium-based. That is why it can preserve familiar browser behaviour and support most Chrome Web Store extensions, although Chrome-specific integrations and some extension behaviours are not guaranteed to work identically.
Can Comet Use Chrome Extensions?
Most Chrome Web Store extensions are supported on desktop according to Perplexity. Users should still review extension permissions and test critical workflows before migrating, because an AI browser plus a high-permission extension creates a larger combined trust surface.
Does Comet Store My Browsing History Online?
Ordinary browsing information is primarily stored locally. However, when an Assistant request explicitly needs page, history, or tab context, Perplexity says the required context can be retained for up to 30 days to support temporary Library and query-history functions.
Is Comet Safe for Banking and Other Sensitive Sites?
Comet includes safe-browsing, local password protections, site permissions, and enterprise controls, but agentic browsing introduces prompt-injection risk. A conservative approach is to block or limit the Assistant on banking, healthcare, payroll, password-management, and administrator sites and avoid broad persistent permissions.
What Is the Difference Between Comet Assistant and Comet Agent?
Assistant is the contextual reading and research companion in the browser, while Agent is the execution layer that can navigate, fill forms, schedule, organise, and complete multi-step workflows. Enterprise documentation presents them as distinct roles even though both appear within the Comet experience.
Is Comet Better Than Chrome?
It depends on the task. Comet is stronger when AI context, cross-tab synthesis, or controlled automation removes several manual steps. Chrome remains a safer default when extension predictability, established enterprise management, sensitive transactions, or a simple one-click workflow matters more than AI assistance.
References
Perplexity. (2026, July 16). Getting Started with Comet.
Perplexity. (2026, March 4). Comet Assistant Privacy & Data Use.
Perplexity. (2026, July 22). Which Perplexity Subscription Plan Is Right for You?.
Perplexity Research. (2025, December 2). BrowseSafe: Understanding and Preventing Prompt Injection Within AI Browser Agents.
Trail of Bits. (2026, February 20). Using Threat Modeling and Prompt Injection to Audit Comet.
Shamsabadi, A. S., Haddadi, H., & Chaikin, A. (2026, June 8). Indirect Prompt Injection Remains a Fundamental Security Challenge for AI. Brave.
Yang, J., Yonack, N., Zyskowski, K., Yarats, D., Ho, J., & Ma, J. (2025). The Adoption and Usage of AI Agents: Early Evidence from Perplexity. arXiv.
Goodwin, D. (2026, March 19). Perplexity’s Comet for iOS Uses Google Search by Default. Search Engine Land.
Xia, Y., Batorski, D., Wertz, E., Mamakos, M., Li, L., & Wojcieszak, M. (2026). AI-Powered Browsers Are Broadly Accurate News Summarizers That Reduce Political Bias and Negative Affect. arXiv.
