OpenAI just moved ChatGPT's agent out of a sandbox and into the browser people already use. Chrome users can now let ChatGPT look at their open tabs and summarize what is on them. In work mode it goes further: it controls the tabs, sending messages, filling in forms, downloading files and tidying tabs into groups.
That sounds like a productivity feature, and it is one. But it also changes who, or what, reads your website when a buyer is deciding between you and someone else.
What ChatGPT can do in Chrome now
There are two levels to the update, per OpenAI's announcement and its help docs for the browser extension.
- Reading. In a normal chat, ChatGPT can see the tabs you have open and pull information from them. You can mention a specific tab, or highlight text and ask about it from a side panel.
- Acting. In work mode, ChatGPT drives the browser itself. OpenAI's own example is telling the browser to open Salesforce and update an account from a set of call notes.
The extension works in Chrome, Edge, Brave, Opera and Vivaldi, and it runs inside the browser profile where you installed it. That last detail matters most. The agent can work on sites where you are already signed in, such as LinkedIn, Gmail or your CRM.
The permission model is per site
By default, ChatGPT asks before it touches each new website, keyed on the host name. The user can allow once, allow for that site, allow all sites, or decline. Allow and block lists live in the app's Computer Use settings, and reading browser history needs a fresh prompt every time.
OpenAI's guidance to users is blunt: treat page content as untrusted context. In other words, the company is telling people the pages its agent reads may be trying to steer it. That is the same problem every answer engine has with web content, now running on a user's own machine.
Why this is a different kind of visitor
Until now, most AI traffic to a company's site came from crawlers. They fetch pages server-side, identify themselves with a user agent, and either respect robots.txt or don't. You could at least see them in your logs and decide what to let in.
An agent inside the buyer's Chrome is not that. It arrives through a real browser with real cookies and the user's own session. For most purposes it is the user. Your robots.txt does not apply to it, and your analytics will mostly record it as a normal visit.
What changes is what it does once it lands. A person skims, scrolls, gets distracted by a hero animation and maybe books a demo. An agent asked to "compare these three vendors on pricing and SSO support" reads the page for those facts, and if it cannot find them quickly, moves on to the next tab.
What the agent needs from your page
If an agent is doing the shortlist, the pages that win are the ones that answer its question plainly. In practice that means:
- Facts in text, not images. Pricing tiers, limits and integrations that live in a screenshot or behind a "contact us" button are invisible or a dead end.
- One page per question a buyer asks. "Does it support SAML?" should land on a page that says yes or no in the first paragraph.
- Forms that work without guesswork. If the agent is told to request a demo, a form with clear labels and no fragile custom widgets is one it can actually complete.
- Comparisons you wrote yourself. If the agent opens a review site or a competitor's comparison page instead of yours, that page frames you.
None of this is new advice for human visitors either. The difference is that an agent won't give a confusing page a second look.
The tab it opens first still comes from an answer
There is one more step that happens before any of this. When a buyer asks ChatGPT to "find me three tools for X and compare them," the agent has to decide which three sites to open. That choice comes from the same search and retrieval layer that powers ChatGPT's answers: the brands it already names, and the pages it already cites.
So browser control doesn't replace answer-engine visibility. It stacks on top of it. First you need to be in the answer, then your page has to survive the agent reading it.
The bigger picture
For the last two years the question for B2B marketers has been "does ChatGPT mention us?" With agents operating inside buyers' browsers, there is a second question: when it opens our site, can it find what it was sent to find?
Both questions have the same starting point: knowing which brands the engines name for your category today, and which pages they are reading to decide.
ClawWorld runs AI visibility work for B2B startups. Start with a free baseline report: 10 buyer questions, who ChatGPT and Claude name instead of you, and the pages they cite.