ChatGPT agent was retired. What now visits your site, and how to let it in

ChatGPT agent was OpenAI’s answer to a simple idea: let ChatGPT use a browser and finish tasks for you. It no longer exists under that name. OpenAI’s help page now says “ChatGPT agent is no longer available”^[1]. The idea did not go away, though. It moved into ChatGPT Work and a cloud browser that still visits websites for people. This post covers what changed, what now reaches your site, and what to set up.

From Operator to ChatGPT agent to ChatGPT Work

OpenAI Operator came first. It launched on 23 January 2025 as a research preview for Pro users in the US: an agent with its own browser, powered by a model OpenAI called Computer-Using Agent, trained to work with “the buttons, menus, and text fields people see on a screen”^[2]. In July 2025 OpenAI merged it into ChatGPT as ChatGPT agent. It ran on its own virtual computer with a visual browser, a text browser and a terminal, and asked permission before “actions of consequence”^[3]. The standalone Operator site was retired.

Atlas, the browser OpenAI launched in October 2025 with agent mode in preview^[4], went the same way. In July 2026 OpenAI said it was shutting Atlas down and spreading its features across the ChatGPT desktop app, a Chrome extension, and “a separate cloud browser” that runs on OpenAI’s servers where its agents complete tasks for users^[5]. OpenAI’s own Atlas launch post now says Atlas “has since been deprecated”.

What visits your site now

The cloud browser is the part that matters to a website. OpenAI describes it as ChatGPT Work’s “own browser on a separate computer in the cloud”, which can read pages, click buttons, fill in forms and act on public and signed-in sites. It is for paid plans, not Free or Go, and ChatGPT decides when to use it^[6]. OpenAI’s examples are ordinary errands: checking a restaurant’s availability, comparing product availability and local stock, finding an appointment, comparing plans on a utility account.

That sits alongside the four identities OpenAI already documents^[8]. They do different jobs, and robots.txt treats them differently:

How to tell a real OpenAI visitor from a fake one

A user agent is just text the visitor sends, and anyone can copy it. Cloudflare put it plainly: user agent headers are “easily spoofed and are therefore insufficient for reliable identification”^[10]. That is why OpenAI publishes an IP list for each crawler, and why its agents sign their requests. OpenAI was already signing every Operator request in May 2025^[10], and Cloudflare listed ChatGPT agent in its first group of “signed agents” that August^[11].

The cloud browser keeps that scheme. Each request carries a Signature header, a Signature-Input header and Signature-Agent set to "https://chatgpt.com", and the public keys are published so your CDN can check them^[7]. Paste a user agent or a set of headers here to see which identity a request claims and which check would prove it:

Letting the cloud browser in

If a bot-protection layer sits in front of your site, it may be blocking people’s tasks without you knowing. OpenAI’s allowlisting page lists the setup for each major provider^[7]:

• Cloudflare: a WAF custom rule on Bot Detection ID 129220581, listed as ChatGPT Operator, set to allow or skip.
• Akamai: a custom bot category containing the Akamai-defined bot “ChatGPT Agent”, set to Allow.
• HUMAN: turn on “ChatGPT Agent” under Known Bots & Crawlers and select Allow.
• Vercel: nothing to do. Vercel recognises the traffic and verifies the signatures itself.

All four verify the signature for you. If your CDN does not, OpenAI’s steps are: check the header matches exactly, quotation marks included; fetch the key; verify the signature under RFC 9421; and only then allow the request. If real traffic is still blocked, a proxy in the middle is often stripping the headers.

Signature-Agent: "https://chatgpt.com"
Signature-Input: <signature parameters>
Signature: <signature>

# Public keys:
# https://chatgpt.com/.well-known/http-message-signatures-directory

What robots.txt still does

robots.txt remains the right tool for the crawlers. Disallowing GPTBot opts your pages out of training; OpenAI recommends allowing OAI-SearchBot if you want to appear in ChatGPT search answers^[8]. OpenAI’s tools have also stopped when told to: in a Cloudflare test, ChatGPT-User fetched the robots file and “stopped crawling when it was disallowed”^[12]. The full reasoning on the training crawler is in should you block GPTBot, and the robots.txt generator writes the file for you.

User-agent: OAI-SearchBot
Allow: /

User-agent: GPTBot
Disallow: /

Making a site that agents can use

OpenAI’s publisher FAQ is specific about how its agent read pages in Atlas: through ARIA, “the same labels and roles that support screen readers”. It recommends following WAI-ARIA best practices, with descriptive roles, labels and states on buttons, menus and forms^[9]. In practice, what helps a screen-reader user helps an agent:

The last item matters more than it looks. When an agent completes a task, the person may never load your page themselves. The GA4 guide covers counting ChatGPT referrals, and AI traffic analytics separates AI visits from the rest. For the bigger question of whether ChatGPT recommends you in the first place, start with the ChatGPT SEO guide.

Questions people ask

Sources