Quick Answer: ChatGPT Atlas means your pages must work for an AI agent acting on a user's behalf, although Atlas itself shut down on 9 August 2026. OpenAI moved the browser agent into the ChatGPT desktop app and Codex, and Perplexity's Comet browser does the same job. A page an agent cannot fetch, read or operate cannot be cited or bought from.
OpenAI launched ChatGPT Atlas on 21 October 2025 as a browser with ChatGPT built in, then deprecated it less than ten months later. The product is gone; the behaviour it introduced is not. OpenAI says it is "building on what we learned from Atlas to support a more capable browser experience in ChatGPT," per its deprecation notice.
This page covers what an agentic browser actually does to a web page, what OpenAI and Perplexity document about how their agents read sites, and what a site owner should change. It does not speculate about ranking factors, because neither company has published any for agent behaviour.
Is ChatGPT Atlas still available
No. Atlas was scheduled to stop working on 9 August 2026, after an announcement planned for 9 July gave users roughly 30 days to export bookmarks and saved pages. OpenAI's announcement page now carries the line "Atlas has since been deprecated."
The capabilities did not disappear. OpenAI states it is moving browser-based agentic work into ChatGPT and Codex, with a new ChatGPT desktop app that supports multiple tabs, downloads, improved navigation and account login, plus a ChatGPT Chrome extension or sidebar for browsing alongside Chrome.
Atlas at a glance
| Question | What OpenAI documents |
|---|---|
| Launched | 21 October 2025, macOS, worldwide, for Free, Plus, Pro and Go users |
| Core features | Ask ChatGPT sidebar on any page, optional browser memories, agent mode |
| Agent mode at launch | Preview for Plus, Pro and Business users |
| Agent limits | Cannot run code in the browser, download files or install extensions; pauses on sensitive sites |
| Deprecated | Scheduled to stop working 9 August 2026 |
| Successor | Browser-based agentic work inside the ChatGPT desktop app and Codex; ChatGPT Chrome extension or sidebar |
Why the shutdown does not change the SEO question
OpenAI says the browser-based agentic capabilities from Atlas moved into ChatGPT and Codex rather than being dropped. A site that was hard for Atlas to use is equally hard for its successor, so the preparation work described below is unchanged by the product decision.
What an agentic browser does with your pages
An agentic browser reads a page on the user's behalf, summarises it, and in agent mode clicks, scrolls, fills forms and completes tasks across sites. OpenAI's Atlas announcement describes ChatGPT "opening tabs and clicking in your browser to complete the task," in its 21 October 2025 post.
Three behaviours a site owner has to plan for
- Reading. The sidebar answers questions about the current page, so the page's text is being interpreted, not just displayed. OpenAI's example is a student asking about lecture slides without switching tabs.
- Remembering. With browser memories on, the agent recalls context from visited sites and brings it back later, such as summarising job postings seen last week. Memories are optional and under the user's control.
- Acting. In agent mode the browser researches, compares, adds items to a cart and books appointments. OpenAI's dinner-party example has the agent finding a grocery store and ordering ingredients from a recipe.
What the agent will not do
OpenAI's safeguards matter for planning. The agent cannot run code in the browser, download files or install extensions, cannot reach other apps or the file system, and pauses on sensitive sites such as financial institutions. A user can also run it logged out, which means your site may be visited by an agent with no session and no cookies.
What Perplexity's Comet does
Comet is Perplexity's Chromium-based browser with the Comet assistant built in, available for Mac, Windows, iOS and Android, per Perplexity's Comet page. Perplexity describes the assistant "actively conducting entire browsing sessions" and executing complete workflows, such as finding another site that ships the same bike faster.
What ChatGPT Atlas changes for SEO
The change is that an agent, not only a crawler or a person, now decides whether your page is usable, and it decides in seconds. Three documented requirements follow from OpenAI's own guidance, and none of them is a ranking factor.
The page must be fetchable by OpenAI's bots
For a site's content to appear in ChatGPT search results, including in the browser, OpenAI's publisher FAQ says you must not block OAI-SearchBot, per the Publishers and Developers FAQ. The same FAQ notes that a disallowed page can still surface as a bare link and title if OpenAI obtains the URL elsewhere, and that a noindex tag is the way to prevent that.
Training is a separate switch. Publishers who want to exclude pages from potential training should disallow GPTBot, and OpenAI states it respected that signal for content acquired through Atlas. The robots.txt rules for each AI bot are the place to set both.
The page must be operable by an agent
OpenAI states that Atlas used ARIA tags, the same labels and roles that support screen readers, to interpret page structure and interactive elements. Its advice to developers is to follow WAI-ARIA practices by adding descriptive roles, labels and states to buttons, menus and forms so the agent recognises what each element does.
That turns accessibility from a compliance item into a commercial one. A checkout button with no accessible name is a button the agent may not find, and a product filter built from unlabelled divs is a filter it cannot use.
The answer must be on the page, not behind a click
An agent summarising a page quotes what is in the rendered HTML at fetch time. Content that arrives only after a scroll event, a tab click or a client-side render is content the summary can miss, which is why structuring content for AI extraction and server-side rendering matter more for agents than they did for classic crawling.
How to prepare a site for browser agents
Prepare for browser agents by making every important page fetchable, server-rendered, labelled and self-describing, in that order. The list below is drawn from what OpenAI and Perplexity document, not from guesses about how their agents rank pages.
- Allow OAI-SearchBot and PerplexityBot in robots.txt and check CDN and bot-mitigation rules, since a 403 from a firewall blocks the fetch before robots.txt is read.
- Server-render the answer. Confirm the page's main claim, price, specification or comparison is present in the initial HTML response.
- Label interactive elements. Give buttons, menus, forms and filters descriptive ARIA roles, labels and states.
- Make product data explicit. Price, availability, variants and shipping should be in the page text and in schema markup an AI can read, so an agent comparing options does not have to infer them.
- Keep the logged-out view complete. An agent may browse logged out, so gate as little as possible and never hide the core answer behind a session.
- Tag inbound traffic. OpenAI appends
utm_source=chatgpt.comto referral URLs from ChatGPT search, so tracking AI-referred sessions separates agent-driven visits from the rest. - Test with the agent. Ask ChatGPT with search enabled, and Comet's assistant, to complete a real task on your site, and fix whatever they fail on.
For e-commerce sites specifically
An agent that can add items to a cart and order them is a buyer that never sees your banner. The product feed and page data that get products into ChatGPT shopping are the same data an agent reads when it compares your listing with another store's, so treat feed quality as a conversion input.
For B2B and content sites
The agent's summary is your first impression. If the page's answer to "does this vendor do X" is buried under a hero image and three paragraphs of positioning, the summary will say so, and optimising for ChatGPT search starts with putting the answer first.
What not to assume about agentic browsers and ranking
Do not assume that either company has published ranking factors for how an agent chooses which sites to visit, because neither has. OpenAI documents access rules and accessibility guidance; Perplexity documents what Comet can do. Anything beyond that is inference.
What is documented
OpenAI states that any public website can appear in ChatGPT search. It also states that OAI-SearchBot access is required for content to be included in summaries, and that ARIA tags improve how the agent operates a site. Perplexity states that Comet is Chromium-based, runs on Mac, Windows, iOS and Android, and supports most Chrome extensions.
What is not documented
Neither company publishes how an agent picks a site to act on, whether accessibility affects that choice, or whether agent visits feed back into search rankings. The honest position is that the same signals that earn a citation, covered in answer engine optimization, are the signals an agent encounters, and that an operable page converts the visit while an inoperable one wastes it.
Where to watch for changes
OpenAI's Atlas release notes carry a last entry dated 10 March 2026, and the deprecation article points to the ChatGPT desktop app for what comes next. Perplexity keeps a separate help centre for Comet. Both are the primary sources to re-read after each major model launch, since agent behaviour changes with the model behind it.