ChatGPT Chrome Extension: Make Your Site Agent-Ready

On October 15, OpenAI shipped a big update to the ChatGPT Chrome extension. ChatGPT can now read across a person's open tabs and, in Work mode, take control of those tabs to fill out forms, send messages, download files and group tabs (Unite.AI). The Chrome Web Store listing reports about 7 million users, and it also runs in Edge, Brave, Opera and Vivaldi.

That changes who is using your website. Some of the "visitors" filling in your signup form or comparing your pricing page next week will be an agent acting for a real, signed-in customer. If your forms are vague, your buttons are unlabeled or your product data only exists as pixels, the agent guesses. Sometimes it guesses wrong, and sometimes it just gives up.

TL;DR

The ChatGPT Chrome extension now acts inside a user's real browser session: it reads tabs, fills forms and clicks through flows on sites the user is signed in to.
Because the agent runs in the person's own browser, robots.txt rules for GPTBot don't decide whether it reaches your pages. Your HTML quality does.
Clear form labels, autocomplete hints, real <button> elements, JSON-LD and (optionally) WebMCP form tools make your site far easier for these agents to use correctly.

What the new ChatGPT Chrome extension actually does

The update splits into two modes. In Chat, ChatGPT looks across the tabs you have open and answers questions about them. OpenAI's own demo compares three sofa listings in three tabs without the user copying a single link (Chrome Geek).

In Work mode, ChatGPT gets control. It can fill out a registration form from an event brief and stop before submitting, move vendor details from several tabs into a comparison sheet, or fill out a new patient form and add the appointment to a calendar. OpenAI says it works "in the background, without ever taking over your active browsing session" (PiunikaWeb). PiunikaWeb reports that the extension is rolling out to free and paid users, with Work mode needing at least a Plus subscription.

The permission model

By default, ChatGPT asks before it uses each new website, keyed to the host. Users pick Allow once, Allow for this site, Allow for all sites or Decline, and can manage allowlists and blocklists in settings (ChatGPT docs). "Allow for all sites" is flagged as the higher-risk option.

This follows OpenAI retiring its standalone Atlas browser, so ChatGPT agent mode now lives in the browser most people already use.

Why this matters for your website

Most AI visibility work so far has been about crawlers like GPTBot and ClaudeBot, which you control with robots.txt (see our robots.txt guide for AI crawlers).

A browser agent is a different animal. It runs inside the customer's own Chrome profile, with their cookies and their login. OpenAI's docs don't describe a separate user agent for it. From your server's point of view, it's your customer's browser loading your pages, so this is not a line you add to robots.txt:

# This controls OpenAI's crawler, not a user's browser agent
User-agent: GPTBot
Disallow: /

So you can't block it like a crawler, and you probably shouldn't: there's a real customer behind it. Your lever is how understandable the page is once the agent arrives.

And the failure modes are boring and familiar. PiunikaWeb lists dynamic websites, unexpected popups and anti-bot checks as the usual things that trip browser agents up. Every one of those is a design choice on your side.

How a browser agent reads your page

OpenAI hasn't published exactly how the extension perceives pages. Browser agents generally lean on the same signals a screen reader uses: the DOM, element roles, accessible names and visible text. So a form that works well with assistive tech usually works well for an agent too.

Compare these two inputs:

<!-- Hard for an agent: no label, placeholder only, generic div button -->
<input type="text" placeholder="Name">
<div class="btn">Next</div>

<!-- Easy for an agent: labeled, typed, autocomplete hint, real button -->
<label for="full-name">Full name</label>
<input id="full-name" name="full_name" type="text" autocomplete="name" required>
<button type="submit">Continue to payment</button>

The second version tells the agent what the field is and what the button does. That's the difference between a first-try success and an email typed into a phone field.

Signal	Helps the agent	Trips the agent up
Form fields	`<label>`, `autocomplete`, correct `type`	Placeholder-only fields, unlabeled inputs
Actions	`<button>` with clear text	Clickable `<div>`s, icon-only buttons with no `aria-label`
Page structure	`<main>`, `<nav>`, one `<h1>`, logical headings	Everything in nested `<div>`s
Product data	JSON-LD `Product` with price and availability	Price only rendered in an image or canvas
Interruptions	Dismissable, labeled dialogs	Surprise popups, CAPTCHA on every step

For product pages, structured data does a lot of the work. Our post on schema markup for AI agents walks through the Product and Offer fields that matter most.

WebMCP: telling agents what your forms do

Labels and roles let an agent infer what a form does. WebMCP lets you say it directly. Chrome opened a WebMCP origin trial in Chrome 149, with an imperative JavaScript API and a declarative API that annotates ordinary HTML forms as tools (PPC Land). A declarative form tool looks roughly like this:

<form toolname="book_demo"
      tooldescription="Book a product demo for a given work email and date">
  <label for="email">Work email</label>
  <input id="email" name="email" type="email" autocomplete="email"
         toolparamdescription="The visitor's work email address">
  <button type="submit">Book demo</button>
</form>

The trade-off is honest: WebMCP is still an origin trial, attribute names can change, and there's no public statement that the ChatGPT extension calls WebMCP tools today. Treat it as a forward bet layered on top of good HTML, not a replacement for it. If you want the full comparison, read WebMCP declarative vs imperative API.

What site owners should do this month

You don't need a rebuild. Start with the pages where an agent is most likely to act for a customer: signup, contact, booking, checkout and pricing.

Label every input. Use a real <label> or aria-label, never a placeholder alone.
Add autocomplete values (name, email, tel, postal-code, cc-name) so agents and browsers know what goes where.
Use real buttons with specific text. "Continue to payment" beats "Next". Give icon buttons an aria-label.
Fix your landmarks and headings: one <h1>, a <main> element and headings that follow the page's logic.
Publish JSON-LD for products, offers, organization and FAQs so facts don't live only in visual layout.
Review interruptions. Delay newsletter popups, keep cookie banners dismissable by a labeled button, and don't put a CAPTCHA in front of low-risk forms.
Test with an agent yourself. Install the extension, open your signup flow and ask ChatGPT to fill it in and stop before submitting. Watch where it hesitates.
Consider WebMCP form tools on your two or three highest-value forms once the basics pass.

Before and after you work through the list, run a free AgentReady scan on your key pages. The Browser Readiness pillar checks the exact signals above: ARIA and landmarks, heading structure, JSON-LD and Open Graph, and WebMCP form tools, so you can see which pages need attention first.

Check your own site

The ChatGPT Chrome extension mostly tests the Browser Readiness pillar of the Agent Readiness Score, which makes up 40% of the total. A low score there usually means unlabeled form fields, missing landmarks, weak heading structure or no structured data, so agents have to guess. A single-page scan is free and takes about a minute. If you also want a clean content map for AI tools, our llms.txt Generator builds one in seconds. Scan your site free.

Key takeaways

The ChatGPT Chrome extension can now read tabs and, in Work mode, fill forms and click through flows in a user's real browser.
robots.txt rules for GPTBot don't control this kind of agent, because it runs in the customer's own session.
Labels, autocomplete, real buttons and landmarks decide whether the agent completes a task correctly.
JSON-LD keeps prices, availability and company facts readable without relying on visual layout.
WebMCP form tools are a sensible next step for high-value forms, but they sit on top of good HTML.

FAQ

What can the ChatGPT Chrome extension do on a website?

In Chat mode it reads open tabs and answers questions about them, such as comparing products. In Work mode it controls tabs to fill out forms, send messages, download files and group tabs. It works on sites where the user is signed in and asks permission per website by default.

Can I block the ChatGPT Chrome extension with robots.txt?

Not in the way you block GPTBot. robots.txt is aimed at crawlers, while the extension acts inside a real user's browser session with their cookies and login. OpenAI's docs don't describe a separate user agent for it. Blocking it would also mean blocking a real customer, so improving your forms and markup is usually the better move.

How do I make my website forms work with AI browser agents?

Give every input a proper label and the right autocomplete value, use real <button> elements with specific text, and keep page structure clear with landmarks and headings. Avoid placeholder-only fields, surprise popups and unnecessary CAPTCHAs. For high-value forms, consider adding WebMCP declarative attributes so agents can see what the form does.

Is ChatGPT agent mode the same as the Chrome extension?

They're closely linked now. After OpenAI retired its Atlas browser, its agent features moved into the ChatGPT browser extension, and Work mode is how ChatGPT agent mode controls tabs in Chrome, Edge, Brave, Opera or Vivaldi. PiunikaWeb reports Work mode needs at least a Plus subscription.

Your next step

Open the form that makes you the most money and check it against the eight steps above. Then see how that page scores on Browser Readiness. Run a free AgentReady scan.

ChatGPT Chrome Extension Can Now Use Your Site. Is It Ready?