The major browsers now ship AI features natively — Microsoft Edge's Copilot integration with page summaries and in-context chat, Chrome's growing set of AI capabilities including page features and assistant-adjacent tools, and the ecosystem of sidebar assistants — and the documented pattern across all of them: the assistant can read the page you're on, summarize it, answer questions about it, and in newer releases act on pages agentically, with the settings to control when. What to enable is a reading-and-acting decision: reading features are conveniences with privacy texture, acting features are the agent-risk class covered earlier in this series — approval gates, scoped powers, outcomes audited.
RechargeMe publishes information, not advice. Details follow browser vendors' published documentation as of early 2026; browser AI ships continuously, so check current settings pages.
What do the browsers document?
Edge: Copilot in the sidebar, per Microsoft's documentation — page summarization, contextual chat about the visible page, compose features, and tab management assistance, tied to the Microsoft account ecosystem. Chrome: Google's documented additions — page summarization and AI features in the browsing surface, Gemini integration points, and the agentic browsing capabilities under development through 2025-2026 with the browser acting on pages under user direction. The general shape everywhere: the page becomes context — the assistant's answer about the article is grounded in the article without your copying it, which is the convenience; the assistant holding page content as its working material is the consideration.
What are the reading features good at?
Three documented strengths. Page Q&A: asking the open page questions — the long article's key claims, the documentation page's relevant setting, the product page's actual warranty terms — without reading the whole thing; grounded in the page, so better than asking a chatbot from memory. Summarization: the tl;dr for triage — deciding whether the long report deserves the full read, with the same caveat as all summaries: it inherits the source's errors and can drop the load-bearing nuance, so the triage summary is a door, not a destination. And comparison assistance: holding several tabs' content and extracting the differences — the pricing pages, the spec sheets — a task browsers were always bad at and assistants do well.
| Feature class | Documented behavior | Enable posture |
|---|---|---|
| Page Q&A / summaries | Grounded in visible page | Convenience — check load-bearing claims |
| Compose / rewrite in page | Drafting in text fields | Watch what fields it touches |
| Tab and comparison help | Cross-page extraction | Strong for shopping/spec work |
| Agentic browsing | Browser acts on pages | Approval gates; scoped tasks |
Related stories: The Claude mobile app: what it adds beyond the browser, per Anthropic's docs · Raycast as an AI launcher: the Mac power tool, documented.
What are the acting features, and their risks?
The newer tier: browser agents that navigate, click, fill, and complete multi-step tasks on pages — documented by both major vendors through 2025-2026 as the agentic frontier moved into the browser. The risk class is the one this series established for agents generally: an actor with page access can be directed by the page — the indirect prompt-injection pattern, where content the agent reads (a page, an email rendered in the browser) contains instructions aimed at the agent — plus the ordinary acting-system risks of wrong purchases, wrong buttons, and compounding steps. The documented controls that matter: approval gates before consequential actions; scoping — the agent asked to research does not need checkout access; and auditing what ran after each agentic session, the outcome-check habit. Browser agents are powerful errand-runners with the largest action radius consumers have handed an AI — enable slowly, watch closely.
What are the privacy settings to review?
The browser-specific layer, per vendors' documentation: what the assistant sends — page content is the working material, so the settings governing browsing-data use for AI features (Chrome's and Edge's privacy pages both describe their arrangements, tied to sync and account settings); when it listens — the hotkey and selection features that can send highlighted text on command, worth knowing so a careless selection doesn't become a disclosure; and the account boundary — browser AI rides the Microsoft or Google account, so the data terms are the ecosystem's, and enterprise versions carry stronger terms than consumer, the recurring pattern of this series. The 30-minute settings audit this series prescribes for every AI product applies with one browser twist: the browser sees everything, so its assistant's permissions deserve the most careful read of any tool on the machine.
Which should a careful user actually enable?
The tiered posture: reading features — page Q&A, summaries, comparison — enabled, with the verification habit (load-bearing claims get checked against the page itself, which is open anyway, making this the cheapest verification in the entire AI toolkit). Compose features — on if you draft in the browser, with the fields-it-touches awareness. Agentic features — off by default in most setups, enabled per-errand when a task genuinely warrants it, gated, scoped, audited. The general rule the whole series supports: adopt the reading layer now, adopt the acting layer deliberately, and re-read the settings page after every major browser release, because the features arrive enabled-in until you decide otherwise.
FAQ
- Does the browser's AI read all my tabs? Per vendor documentation, page features work with the page in front of you; agentic features work with what you direct them to. Settings govern browsing-data use — read them for your account type.
- Are browser agents safe? They're the agent risk class: approval gates, scoped tasks, outcome audits — and the prompt-injection consideration that a page can carry instructions aimed at the agent. Enable per-task, not globally.
- Which browser AI features should I use? Start with page Q&A and summaries — grounded, convenient, cheap to verify against the open page — and hold agentic browsing for tasks that genuinely need it, gated.