ChatGPT Work Can Operate Websites: How to Make Yours Agent-Ready – Article and Video 64

ChatGPT Can Now Operate Signed-In Websites: How to Make Your Site Agent-Ready

Author: Kevin C. Roy  ·  Published: September 15, 2026

What Changed?

OpenAI has expanded ChatGPT Work’s cloud-browser capabilities.

The browser can operate supported public and authenticated websites and perform actions such as:

• Reading webpages
• Clicking buttons
• Entering information into forms
• Navigating multi-step workflows
• Continuing tasks after the user signs in
• Maintaining a signed-in session for later supported tasks. (OpenAI Help Center)

Authentication is handled through a secure sign-in experience. OpenAI says the model does not see or store the user’s username and password.

For consequential actions—including certain bookings and payments—ChatGPT is designed to stop and request confirmation before completing the action. (OpenAI Help Center)

OpenAI’s August 25 release notes describe examples including finding a DMV appointment, working through utility-plan signup and handling other signed-in workflows. (OpenAI Help Center)

This represents an important shift in how AI can participate in a customer journey.

From Recommendation Engine to Action Engine

The early generative-search funnel looked roughly like this:

Question → AI answer → recommendation → website visit

Increasingly, another stage is possible:

Question → recommendation → delegated action

A user might ask an AI assistant to:

• Find a suitable service provider
• Request an appointment
• Find an available product
• Select an appropriate option
• Begin registration
• Complete a form
• Navigate a customer portal

That means the website is no longer necessarily being operated directly by the human customer.

Software may increasingly be acting on the customer’s behalf.

What Does This Mean for Website Optimization?

This is not a disclosed search-ranking update.

Clear buttons and forms are not suddenly documented ChatGPT or Google ranking factors.

The opportunity happens farther down the funnel.

Once an AI system identifies a business or product, can it successfully understand and operate the website?

That creates a useful new concept:

Agent-Ready Conversion Design

An agent-ready website makes its important commercial workflows understandable and operable by both people and automated assistants.

The objective is not to build a separate experience for AI.

The objective is to remove unnecessary ambiguity.

A useful rule is:

If a user must infer what a control means, an agent may have to infer it too.

The Four-Part Agent-Ready Framework

For important conversion workflows, test four stages:

1. Find

The first step is correct destination selection.

For a service company, that may mean distinguishing between:

• Repair
• Installation
• Maintenance
• Emergency service

For ecommerce, it may mean finding:

• The correct product
• Model
• Size
• Variant
• Configuration

Clear architecture and naming reduce ambiguity.

2. Understand

Once the correct destination is found, the agent must understand the available choices.

Useful information includes:

• Clear pricing
• Availability
• Service descriptions
• Product specifications
• Size and variant information
• Appointment types
• Geographic limitations
• Eligibility requirements

This information should be stated explicitly rather than requiring unnecessary inference.

3. Complete

Then test whether the workflow itself is understandable.

Good conversion paths generally use:

• Descriptive button labels
• Properly labeled form fields
• Clear required-field indicators
• Stable form state
• Predictable navigation
• Understandable validation
• Useful error messages
• Accessible semantic controls

For example:

A button labeled “Request HVAC Quote” provides more context than a generic button labeled “Continue.”

The objective is not to optimize copy for robots.

It is to communicate intent clearly.

4. Confirm

The final success state should be unmistakable.

Examples include:

• “Your appointment has been scheduled.”
• “Your quote request was successfully submitted.”
• “Your registration is complete.”
• “Your order was placed successfully.”

Avoid relying entirely on visual changes or ambiguous redirects.

The site should clearly communicate that the requested action occurred.

Five Tasks Every Transactional Site Should Test

Choose five commercially important journeys.

Lead-Generation Websites

Consider testing:

1. Find the correct service.
2. Request a quote.
3. Schedule an appointment.
4. Select the correct location.
5. Submit a lead form.

Ecommerce Websites

Consider:

1. Find the requested product.
2. Select the correct model or variant.
3. Determine whether it is available.
4. Add the correct configuration to the cart.
5. Navigate the checkout process up to the appropriate confirmation boundary.

The exact tasks will vary by business.

What matters is testing actual customer intent rather than arbitrary pages.

A Simple Agent-Readiness Scorecard

Record the result of each test.

This turns agent usability into a diagnostic exercise rather than speculation.

Don’t Ignore Bot Protection

There is also an infrastructure consideration.

OpenAI says some websites may block ChatGPT Work’s cloud browser because security systems restrict automated browser agents. (OpenAI Help Center)

OpenAI provides a mechanism for website operators to verify legitimate ChatGPT cloud-browser requests using signed HTTP messages. This allows CDNs and firewalls to recognize legitimate ChatGPT traffic without simply disabling security controls. (OpenAI Help Center)

For technical teams, review:

• CDN configuration
• WAF rules
• Automated-bot controls
• Rate limiting
• Session behavior
• Authentication workflows

The recommendation is not:

Allow every bot.

It is:

Know whether your security stack is preventing legitimate delegated customer activity.

What Not to Do

This update does not justify rebuilding a website around AI agents.

Businesses do not need:

• A second AI-only website
• Random new schema because someone labels it “agent optimization”
• Reduced security
• Completely separate conversion paths

Those responses would be premature.

Instead, concentrate on well-established website principles:

Clarity.
Consistency.
Accessibility.
Predictable workflows.
Explicit success states.

Those improvements benefit humans today and make the website easier for software agents to operate as delegated browsing becomes more common.

The GreenBanana Agent-Ready Audit

For transactional clients, I would run the audit like this:

Step 1: Identify the Five Money Actions

Choose the five website actions most closely connected to revenue.

Step 2: Write Real Customer Requests

Use natural-language prompts.

For example:

“Find the correct service for a broken AC unit and request an appointment for next week.”

Or:

“Find the waterproof version of this boot in men’s size 10 and add it to the cart.”

Step 3: Run the Tasks

Use ChatGPT Work and other capable agentic browsers where appropriate.

Step 4: Record the First Failure

Do not simply mark the test “failed.”

Record precisely where:

• Wrong page selected
• Choice misunderstood
• Form field unclear
• Navigation failed
• Agent blocked
• Success not recognized

Step 5: Fix the Ambiguity

Prioritize problems that would also make the experience better for real customers.

Then retest.

A Better Way to Think About AI Conversion

AI visibility metrics remain important.

Businesses should still measure:

• Citation visibility
• Recommendations
• Brand mentions
• Referral traffic

But agentic browsing adds another dimension:

Can the recommended company actually be acted upon?

A useful future KPI may therefore become:

Agent Task-Completion Rate

For example:

8 of 10 representative customer tasks were successfully completed by the agent.

Eventually, businesses may evaluate both:

AI Recommendation Share and Agent Task-Completion Rate

because the combination gets much closer to actual commercial opportunity than citations alone.

Key Takeaway

ChatGPT is becoming capable of doing more than identifying the right website.

It can increasingly interact with the website on the user’s behalf.

The sensible response is not fear and it is not another layer of marketing jargon.

It is to make important customer workflows extremely clear.

For each conversion path, ask:

• Can it find it?
• Can it understand it?
• Can it complete it?
• Can it confirm it?

If the answer is yes, the website is becoming much better prepared for the next phase of AI-driven customer behavior.

Add the Watch the Breakdown section once you have the real YouTube video ID.

Watch the Breakdown