Google’s Auto Browse is an early look at a version of the web where an AI agent does the clicking for you. In Chrome, the feature lets Gemini move through tabs, visit sites, choose options, and pause when a task needs human approval.
The idea is straightforward: ask for a real-world browser chore, then watch the agent try to complete it. The test results were more complicated. Auto Browse could navigate, search, and act across several steps, but it also showed how easily a technically correct action can still miss what a person actually meant.
What Auto Browse Is Trying To Change
Google released Auto Browse this week to US users who subscribe to its AI Pro and AI Ultra plans. The feature is available through the Gemini chatbot inside the Chrome sidebar, reached from the small sparkle icon in the browser.
Google’s examples for the tool included reserving tickets, shopping for clothes, and planning vacations. Those are not narrow demo tasks. They are everyday web jobs that often involve preference, judgment, price, timing, and, in some cases, payment details.
That is why Auto Browse matters. It is not just another chatbot answer box. It attempts to operate the browser itself, turning a written request into a chain of clicks. If users learn to trust that workflow, the web could start to feel less like a place people personally navigate and more like a place software agents negotiate on their behalf.
For now, access can be uneven. In the test described by WIRED, the first attempt made Gemini appear to describe browser control without actually taking it. The bot said, “You’ll see me opening and closing tabs,” and claimed it had “taken over the navigation,” but no clicks followed. After logging out of the Google account and refreshing the browser, the Auto Browse option appeared as an opt-in pop-up.
Users can also check settings for the “Let Chrome browse for you” toggle. Once activated, direct requests sent through the Gemini sidebar can trigger Auto Browser to begin clicking.
The Ticket Test Showed Both Promise And Fragility
The first meaningful test involved buying two tickets to the SF symphony. The request was specific: avoid orchestra seating, avoid the cheapest seats, choose two seats next to an aisle, and book for that night.
Auto Browse began by using Gemini 3 to outline a plan in the sidebar. Then it moved through the web task. It went to the correct website, selected the right performance, examined multiple seating sections, and kept a log of its steps. In broad terms, that is exactly the sort of multistep behavior an AI browser agent needs to perform.
After a couple of minutes, the bot stopped and asked the user to take over at the Order Now button. That handoff matters. Google has safeguards for sensitive actions, including buying things or posting on social media, that require user approval before continuing.
At first glance, the task looked complete. But the selected seats revealed the central problem. Auto Browse chose $185 seats in the side box. They were by an aisle and outside the orchestra section, just as requested. They were also in separate rows.
That failure was not a random navigation error. It was a judgment error. A human buying two tickets for a date would usually understand that the seats should be next to each other unless told otherwise. The prompt did not explicitly say that, and the agent followed the written requirements too narrowly.
The result was more work, not less. The user took over and bought two cheaper tickets that were actually next to each other.
Shopping Was Easier, But Not Smarter
The second test asked Auto Browse to find a leather jacket on Depop that could fit an XL male, add three options to the cart, and explain the case for each fashion choice.
On this task, Auto Browse performed more smoothly. It visited Depop, searched for “men’s leather jackets,” applied the XL sizing filter, placed the first three jackets from the results into the cart, and wrote a paragraph about each option.
That is useful in a limited way. It reduced typing and handled some repetitive web actions. One jacket was $40 and looked decent enough in the test.
But the task also exposed a different weakness: selection quality. Adding the first three search results does not demonstrate much taste, comparison, or curation. The agent completed the visible steps, yet it did not appear to make a meaningful qualitative judgment about variety or style.
That distinction is important for AI agents. A browser task is rarely only about moving through buttons and filters. People often want the tool to understand the purpose behind the request, compare tradeoffs, and avoid shallow choices that technically satisfy the instruction.
Why Caution Still Matters
Auto Browse includes visible warnings while it runs. The sidebar tells users: “Use Gemini carefully and take control if needed,” and “You are responsible for Gemini’s actions during tasks.” Those warnings fit the current state of the product.
There are several reasons to stay cautious:
- Generative AI can make mistakes. Google’s Gemini chatbot already reminds users of that risk.
- Browser control raises the stakes. A wrong answer is one thing; a wrong click during checkout is another.
- Prompt injection attacks are a concern. Malicious websites can try to redirect an AI tool away from the user’s intended task.
- Outside researchers have not fully examined Auto Browse’s potential vulnerabilities. The risks may resemble those of other AI tools that can control a computer.
- Purchases deserve extra scrutiny. Even with approval safeguards, users must decide whether they are comfortable involving an agent around credit card activity and financial information.
The early lesson is not that Auto Browse is useless. It can already handle pieces of the browsing process, and it appears more capable at multistep tasks than similar agent tools tested last year. The issue is trust.
For an AI agent to become a dependable web assistant, it must do more than click through pages. It must understand ordinary human expectations, know when a request is under-specified, and stop before turning a small ambiguity into a practical problem. Auto Browse shows that future beginning to arrive in Chrome, but it also shows why people still need to keep one hand on the browser.