Start with the comparison
Imagine looking for a sofa. Does it fit through the door, what is the material, when can it arrive, and what are the return conditions? That comparison is among the examples for ChatGPT’s new browser connection. Bringing information from open pages into one conversation could reduce copying between tabs. It also gives the task a useful endpoint: comparable options with missing details still visible. The browser need not take over the whole purchase for that step to have value.[1]
The connection’s Chat and Work modes change the scope of that small job. Chat works with page content; Work can fill forms, download files and organize tabs. Naming the transition from reading to acting is a useful start. In the sofa example, the assistance I want is collection of dimensions and conditions. Moving into a delivery form, entering an address or placing an order is another task, with a separate stopping point.[1]
Before any of that convenience, setup requires choosing a browser in the desktop app’s Computer Use settings and installing an extension. Other Chromium browsers are supported alongside Chrome, although Opera lacks the side panel. Access can vary by account or workspace. The first practical question therefore comes before a smooth demonstration: can the person connect their browser and bring an open retailer tab into the conversation? That is the initial setup job, rather than a claim that a user test has already succeeded.[1]
When another site opens
Then another retailer’s page enters the task. The default behavior asks for access to a new domain, with temporary or site-specific permissions available. Particular sites can also be blocked. That offers a bounded route for someone who wants only this comparison. Repeated prompts could add effort, while the same pauses could help the person notice which retailer is being opened. The useful comparison is the work left after approvals and corrections, set against the old copying workflow.[1]
Browser-history access requires approval each time. There is also no separate comprehensive action log; text, screenshots and summaries brought into the conversation follow existing data settings. For this task, I would want an output that makes it easy to revisit which product dimension or return condition is being compared. A polished recommendation alone is less useful for that purpose. A smooth list that substitutes a guess for a missing dimension cannot make the doorway wider.[1]
What matters after the comparison is the work still on the person’s desk: can missing information be spotted, can a mistaken product detail be corrected, and does the task stop at the requested point? Those are not demonstrated user outcomes yet. The convenience I would look for after the announcement is a readable comparison from the pages already open, with a comfortable stop before ordering. That is a concrete enough job to judge without turning every browser tab into a promise of a new routine.[1]