ChatGPT Atlas is retired. OpenAI's agent now browses from three places, and its crawler list names none of them.

On 9 July 2026, 9to5Mac's Zac Hall reported that OpenAI was ending ChatGPT Atlas, the browser it had launched less than a year earlier.

On 24 October, OpenAI's own help centre said it plainly. Atlas has been retired and no longer receives updates.

Google autocomplete has not caught up. Checked on 6 November, it still finishes "chatgpt atlas" with "browser", "windows" and "download".

So people are still looking for a product that is gone.

Here is the complication for a business. Agentic browsing did not die with Atlas. It moved. And in moving, it went somewhere OpenAI has documented for users, and for website owners in only one of three places.

Step 1: What happened, from the sources

The 9to5Mac report quotes OpenAI's James Sun on X: "we are going to be sunsetting Atlas". He gave 8/9 as the targeted deprecation date.

The same report says the browsing moved into a new ChatGPT desktop app, launched that day, plus a ChatGPT extension for Chrome.

The Atlas retirement article in OpenAI's help centre showed as updated a day earlier when I read it on 24 October. It told users bookmarks, tabs, history and signed-in sessions do not transfer, and to sign in to websites again in a new browser.

That last line matters to anyone running a site with logins. A user who did tasks through Atlas is now arriving from a different browser with no saved session.

Step 2: Where ChatGPT browses now

OpenAI's help article on the built-in browser in the ChatGPT desktop app, read on 24 October, names three surfaces.

The built-in browser. It runs inside the desktop app on macOS and Windows, with its own browser state. The user and ChatGPT see the same page. ChatGPT can work across tabs, download files and wait while the user signs in.

The Chrome extension. It uses the user's existing Chrome profile, cookies, signed-in sessions, open tabs and extensions. It works with ChatGPT on the web, without the desktop app.

The cloud browser. The help page says it runs remotely in ChatGPT and can continue delegated web tasks in the background.

Two of those three run on the user's own machine. One runs on OpenAI's.

Step 3: What OpenAI has told website owners

Less than you would want.

OpenAI's crawler documentation, read on 6 November, lists four user agents. OAI-SearchBot for search. OAI-AdsBot for checking ad landing pages. GPTBot for training. ChatGPT-User for certain actions a user triggers inside ChatGPT, where the page warns that robots.txt rules may not apply.

None of the four is described as the desktop browser, the Chrome extension or the cloud browser.

So here is what I can say, and where it stops.

A task run through the Chrome extension happens inside the user's own Chrome. Nothing OpenAI has published says it announces itself differently from that person browsing. The built-in browser is a full browser on the user's machine, and the help page does not describe any marker either. The cloud browser runs on OpenAI's side, and the crawler page does not say what it sends.

One of the three is documented, on a different page. OpenAI's Cloud browser allowlisting article, read on 6 November 2026 and marked as updated 17 days earlier, says the Cloud browser signs its outbound requests using Web Bot Auth and the HTTP Message Signatures standard, RFC 9421, with a Signature-Agent header set to "https://chatgpt.com".

It publishes the public keys too, and says Cloudflare, Akamai, HUMAN and Vercel already recognise this traffic under their existing ChatGPT agent entries. It also says that at launch the Cloud browser cannot sign in to websites or complete payments.

So the cloud browser is a question for your CDN, and it has an answer. A valid signature can be checked against a key OpenAI publishes, which is stronger than a name in a log.

The other two stay dark. I have not seen a primary source on how the desktop browser or the Chrome extension identifies itself. If you read a vendor report that claims to count ChatGPT agent visits precisely, ask which of the three surfaces it can see, and how it knows.

We covered the documented bots one by one in GPTBot, OAI-SearchBot and ChatGPT-User. That post still holds. This is the part it could not cover: OpenAI wrote the cloud browser down on a separate page, and the other two nowhere I can find.

Step 4: Do not block what you cannot see

The instinct after a post like this is to reach for robots.txt.

Do not.

robots.txt is a request to crawlers. A person browsing with an assistant in their own Chrome is not a crawler. OpenAI already says ChatGPT-User, the one user-triggered agent it does name, may not follow robots.txt rules because a user started the action.

And blocking the visible part does not stop the invisible part. It only stops the search bot, which is the one that gets you named in ChatGPT answers.

The contrarian line: the agent is a customer.

Someone asked ChatGPT to book, compare or buy, and it came to your site to do it. Make that work.

Step 5: Build pages an agent can use

The good news is that none of this is new advice.

An agent in a real browser reads a page through the same structure a screen reader uses: button labels, form field names, headings, link text. When Atlas launched, OpenAI's announcement, as quoted by Simon Willison on 21 October 2025, said site owners could add ARIA tags to improve how the agent works on their sites.

That held for Atlas. Nothing in OpenAI's current help pages says it stopped mattering, and the built-in browser is still a browser.

So check three things.

Every button and form field has a real label. "Book a table", not an icon with no name.

Key facts sit in text. Prices, hours, availability and policies, not baked into an image.

Sign-in works without tricks. The help page tells users to enter credentials only in the browser, never in the chat. If your login breaks in a fresh browser with no saved state, a user coming from Atlas hits that wall first. VALORAE Media makes the same point from the agency side in never give your agency your password: access should go through the front door.

Lighthouse now has an audit aimed at this, which we walked through in Lighthouse's agentic browsing audits. It is the cheapest check you can run today.

Step 6: The annotation hook, and why I would wait

One line in OpenAI's help page is new.

It says some websites customize annotations for their content. A site can let a user select a whole item, like a product card, and can pass extra context with a selection, such as an item's name, identifier or properties.

That is a hook for sites to hand structured context to ChatGPT inside the browser.

But the help page does not link to developer documentation for it, and I could not find any. I am not going to tell a business to build against an interface I cannot read.

Watch it. Do not spend on it yet.

Step 7: What NovaTechRay changes, and what it does not

At NovaTechRay we do not change our robots.txt advice. Allow OAI-SearchBot if you want to appear in ChatGPT search. Decide GPTBot on its own merits.

We do not report agent visits as a number, because the sources do not support one.

We do keep checking the boring things: labelled controls, text over images, a sign-in that works cold. Those served people before any of this, and they are the only part of agent readiness with evidence behind it.

If OpenAI documents how the other two browsers identify themselves, we will update this post and say what changed.

The short version

• Atlas is retired. OpenAI's help centre says so.
• ChatGPT browses from a desktop app, a Chrome extension and a cloud browser.
• OpenAI's crawler list names none of the three.
• Verify the cloud browser by its signature, at your CDN.
• Do not trust a precise count of ChatGPT agent visits.
• Do not block in robots.txt what you cannot identify.
• Label every control, keep key facts in text, make sign-in work cold.
• Wait for documentation before building for custom annotations.

This is the kind of question we answer at novatechray.com.