- An AI browser puts the AI agent inside the browsing session itself (reading the live page, navigating, filling forms) rather than in a separate chat window that fetches pages into its own context. That's a structural shift in where retrieval and citation actually happen, not just a new interface for an old mechanism.
- The category launched and grew fast through the second half of 2025 and into 2026: Comet (Perplexity, July 2025, free by March 2026), Atlas (OpenAI, October 2025), and AI browsing modes added to Edge (Microsoft, July 2025) and Chrome (Google, September 2025).
- Reported usage figures for these products vary enormously by source and aren't independently reconciled against each other or against the companies' own disclosures, so any specific user-count or traffic-share number should be treated as directional at best.
- A widely-repeated claim that Comet holds 48.12% of agentic web traffic, up 7,851% year-on-year, traces to a cited report (Human Security) that couldn't be independently confirmed in researching this piece.
- Security researchers flag a real, documented risk specific to this category: prompt injection from a malicious site can hijack an AI browser's agentic actions, a risk that doesn't exist in the same form for a chatbot that only reads pages into a separate conversation rather than acting on them directly.
AI browsers are a genuinely new product category, distinct from a chatbot with a web-search feature bolted on, and the mechanics of how they retrieve and present information differ enough to matter for anyone thinking about AI-citation strategy.
What actually makes an AI browser different from ChatGPT with browsing enabled?
The structural difference is where the AI agent sits. A chatbot with web search fetches pages into a separate conversation window; the browsing happens on the model's side, invisible to the user beyond the citations that come back. An AI browser puts the agent inside the actual browsing session: it can read the live page a person is looking at, navigate between pages, fill in forms, and answer questions grounded in whatever's actually on screen, rather than a fetched copy retrieved into a separate chat context. That's a meaningfully different retrieval model, and it means citation, in the traditional sense of a numbered reference in a generated answer, isn't necessarily the primary interaction at all; the agent might just act directly on the page rather than citing it back to the user.
What actually launched, and when?
The timeline moved fast. Perplexity's Comet began as a $200/month exclusive product in July 2025 and had gone completely free across all platforms by March 2026. OpenAI's ChatGPT Atlas launched on macOS in October 2025. Microsoft added a Copilot Mode directly into Edge in July 2025, and Google integrated Gemini into Chrome that September, meaning all four major AI labs had some form of browser-embedded AI agent live within roughly a fourteen-month window.
How much usage do these products actually have?
This is where the available data gets genuinely unreliable. One low-authority source reports Comet's monthly active users at 18 million in Q1 2026, with 2.1 million Pro subscribers, address-bar AI searches hitting 298 million in the same quarter, and agentic tasks growing 860% year-over-year to 14 million. A separate source puts ChatGPT Atlas at roughly 11 million monthly active users by Q1 2026, with 175% quarter-over-quarter growth. Neither figure is independently reconciled against the companies' own official disclosures or against each other, and both come from sources with limited independent authority behind them. Treat the category's overall growth trajectory as real and fast, and any specific number as a rough order of magnitude rather than a confirmed figure.
Is there real data on agentic traffic share specifically?
One widely-repeated claim, attributed to a Human Security report, states that Perplexity Comet holds 48.12% of all tracked agentic web traffic, up 7,851% year-on-year. That specific figure could not be independently confirmed against the original report during research for this piece, which is worth stating plainly rather than repeating the number as settled: it may well be directionally accurate given Comet's early-mover position in the category, but the precise percentage shouldn't be treated as verified.
What's the actual new risk this category introduces?
A real, documented one, distinct from anything in the chatbot-citation discussion elsewhere on this site: prompt injection. Because an AI browser can take agentic actions (filling forms, clicking through flows, potentially handling logged-in sessions) rather than just reading and reporting back, a malicious website can embed hidden instructions designed to hijack the agent's actions directly, a risk category that doesn't exist in the same form for a chatbot that only fetches page content into a read-only context. Security researchers have flagged this specifically as a reason to avoid letting an AI browser agent operate on pages tied to sensitive accounts.
What does this mean for AI-visibility strategy?
The underlying content fundamentals (crawler access, extractable structure, accurate entity data) still matter, since an AI browser agent still has to be able to read and parse a page correctly regardless of interface. What's genuinely new is that "getting cited" may increasingly mean "getting acted on directly" rather than "appearing as a numbered reference in generated text," a distinction worth watching as this category matures rather than something settled enough yet to build a confident playbook around.
Frequently asked questions
What is an AI browser, and how is it different from a chatbot with web search?
An AI browser is a full web browser (not a separate chat window) with an AI agent built directly into the browsing experience itself: it can read the page you're on, navigate, fill forms, and answer questions using the live page content you're actually viewing, rather than a chatbot fetching pages into a separate conversation. That puts the citation and retrieval layer inside the browsing session itself, not in a parallel interface.
Which AI browsers actually launched, and when?
Perplexity's Comet started as a $200/month exclusive in July 2025 and became completely free across all platforms by March 2026. OpenAI's ChatGPT Atlas launched on macOS in October 2025. Microsoft added a Copilot Mode to Edge in July 2025, and Google integrated Gemini into Chrome that September.
How much usage do these AI browsers actually have?
Reported figures vary widely and aren't independently reconciled across sources. One low-authority source reports Comet's Q1 2026 monthly active users at 18 million with 2.1 million Pro subscribers; a different source puts ChatGPT Atlas near 11 million monthly actives by Q1 2026 with 175% quarter-over-quarter growth. Treat both as directional rather than confirmed.
Is there real data on how much agentic web traffic these browsers generate?
One claim, attributed to a Human Security report, states that Perplexity Comet holds 48.12% of all tracked agentic web traffic, up 7,851% year-on-year. That figure could not be independently confirmed against the original report during research for this piece, so treat it as an unverified, high-uncertainty claim rather than settled data.
Sources referenced
For how the fetched-versus-cited distinction already works in the chatbot context this category is extending, see what does it mean to be cited by AI.
