ChatGPT Atlas Browser Discoverability Specification

TL;DR

• Atlas ships an unmodified Chrome user-agent (Mozilla/5.0 ... Chrome/141.0.0.0 Safari/537.36) and has no dedicated robots.txt token, so Atlas-specific blocking via UA filtering is unreliable (OpenAI, 2025).
• Citation eligibility is governed by OpenAI's three documented crawlers — GPTBot, OAI-SearchBot, and ChatGPT-User — not by Atlas itself (OpenAI Crawler Docs, 2026).
• Atlas's agent mode interprets pages through the accessibility tree (ARIA roles, labels, headings); discoverability for agent actions tracks WAI-ARIA conformance more than traditional SEO (Simon Willison, 2025).
• Robots.txt opt-outs apply to OAI-SearchBot for search inclusion and to GPTBot for training; user-initiated ChatGPT-User fetches may bypass robots rules per OpenAI's December 9, 2025 documentation update (ppc.land, 2025).
• Discoverability checklist: allow OAI-SearchBot, expose stable URLs without JS-only routing, render semantic HTML server-side, publish JSON-LD schema, and monitor referer headers and IP ranges from OpenAI's published lists.

Definition

The ChatGPT Atlas Browser Discoverability Specification defines the technical conditions under which a website's content is reachable, parsable, and citation-eligible inside OpenAI's Atlas browser. ChatGPT Atlas is a Chromium-based desktop browser launched by OpenAI on October 21, 2025 (OpenAI, 2025) that ships ChatGPT as an in-browser sidebar and an autonomous agent mode capable of acting on the user's behalf.

Unlike traditional crawlers, Atlas does not identify itself with a unique bot token. Its default network traffic uses the identical Chrome-on-macOS user-agent string and TLS fingerprint as the underlying Chromium build (HUMAN Security, 2025). As a result, "discoverability for Atlas" is a layered concept: the user inside Atlas is a human visitor, while the citation pipeline that powers Atlas's sidebar answers depends on three upstream OpenAI crawlers — GPTBot, OAI-SearchBot, and ChatGPT-User — that do announce themselves.

This specification defines the four discoverability channels (Atlas browser fetches, agent-mode fetches, search-pipeline fetches via OAI-SearchBot, and user-action fetches via ChatGPT-User), enumerates the user-agent strings for each, lists the robots.txt directives OpenAI honors per channel, and provides a paste-ready conformance checklist.

Why this matters

Atlas reframes discoverability for content owners. Traditional SEO assumed a binary — Googlebot fetches your page, ranks it, and surfaces a SERP. Atlas introduces three orthogonal access surfaces:

1. Human-in-Atlas browsing. A real user clicks through, but the Atlas sidebar may still summarize, paraphrase, or extract content without a click on links inside the summary.
2. Agent-mode actions. Atlas's agent autonomously navigates pages on the user's behalf, using ARIA roles to identify form fields, buttons, and content blocks (Simon Willison, 2025).
3. Citation-pipeline ingestion. When in-Atlas ChatGPT generates an answer that cites your site, the underlying retrieval is performed by OAI-SearchBot or ChatGPT-User, not by Atlas itself.

Misunderstanding these layers produces two common failure modes. First, sites that block "anything that smells like a bot" by aggressively filtering Chrome user-agents from non-residential IPs may inadvertently lose Atlas users. Second, sites that allow only a static UA list can lose OAI-SearchBot ingestion when OpenAI rotates its crawler suffix — which OpenAI did on December 9, 2025, when the company also clarified that OAI-SearchBot is no longer used for training (ppc.land, 2025).

Practitioners also report that referrer attribution through Atlas is unstable because in-Atlas summaries often do not generate a click (SALT.agency, 2025). Reliable measurement therefore depends on log-level signals from the upstream crawlers rather than on traditional analytics.

How it works

Atlas discoverability splits across four fetch channels. Each channel uses a distinct user-agent and obeys (or ignores) robots.txt differently. The table below summarizes the contract.

Source: OpenAI Crawler Documentation, revised December 9, 2025.

Verification rules. Atlas browser traffic cannot be cryptographically verified because it spoofs a stock Chrome UA (HUMAN Security, 2025). Verification is only meaningful for the three upstream bots:

1. Confirm the user-agent suffix matches the documented strings.
2. Resolve the requesting IP via reverse-forward DNS to a *.openai.com host, where supported.
3. Cross-check against OpenAI's published IP allow-lists at openai.com/gptbot.json, openai.com/searchbot.json, and openai.com/chatgpt-user.json.

Fetch semantics. Atlas's in-browser ChatGPT sidebar typically reads the rendered DOM of the active tab rather than issuing a fresh fetch from your origin. Citations in agent answers are usually grounded by an OAI-SearchBot lookup that runs in OpenAI's backend. As a result, even a perfectly Atlas-friendly page will not be cited unless OAI-SearchBot can also reach and parse it.

Rendering pipeline. Atlas runs on Chromium 141 at launch, so it executes JavaScript identically to mainline Chrome. OAI-SearchBot, by contrast, has historically had limited or delayed JS rendering. Server-side rendering or static generation for primary content is therefore required if you want the same content to be visible to both Atlas users and the citation pipeline.

Practical application

The following implementation steps are paste-ready for most Linux/Apache/Nginx stacks.

1. Allow OpenAI's crawlers in robots.txt. Add explicit allow rules for OAI-SearchBot (search inclusion) and ChatGPT-User (user-initiated fetches), and decide on GPTBot based on your training-opt-in policy.

User-agent: OAI-SearchBot

Allow: /

User-agent: ChatGPT-User

Allow: /

User-agent: GPTBot

Disallow: /private/

1. Do not UA-block Chrome on macOS. Atlas presents as Chrome 141 on macOS. UA-blocking that signature would exclude Atlas users alongside legitimate human Chrome users. If you must rate-limit, use behavioral signals (request cadence, missing referer plus missing cookies, absent fingerprinting JS) rather than UA strings.
2. Render primary content server-side. Critical content — title, primary headings, body paragraphs, schema — should be present in the initial HTML response so that OAI-SearchBot can parse it without executing JavaScript. Hydrate interactive widgets after.
3. Expose ARIA roles for agent mode. Apply WAI-ARIA Authoring Practices to interactive elements. Atlas agent mode uses the accessibility tree (the same surface used by screen readers) to identify buttons, menus, and form fields (Simon Willison, 2025). Missing or generic ARIA reduces agent reliability.
4. Stabilize URLs. Avoid hash-routed (/#/page) and JS-only routes for indexable content. OAI-SearchBot requires server-resolvable URLs; agent navigation works better with bookmarkable URLs.
5. Add structured data. Publish JSON-LD Article, FAQPage, HowTo, or Product markup. Both OAI-SearchBot and Atlas's sidebar leverage structured data to surface citations more confidently.
6. Log and segment all four UA channels. Tag log lines with oai-searchbot, chatgpt-user, gptbot, and a separate atlas-suspect bucket for Chrome 141 on macOS during business hours.
7. Validate citation eligibility. Periodically query ChatGPT and Atlas for branded queries and confirm that citations resolve to canonical URLs, not paginated or AMP variants.

Common mistakes

• Treating Atlas as a crawler. Atlas is a browser, not a bot. Adding User-agent: Atlas rules to robots.txt does nothing because that token does not exist in the spec. The correct controls live at the upstream crawler layer (OAI-SearchBot, GPTBot).
• Soft-404 on JS-only routes. Pages that return a shell HTML and rely on client-side hydration to populate content frequently appear empty to OAI-SearchBot, eliminating citation eligibility.
• Aggressive Cloudflare or Akamai bot challenges. Default "AI bot" managed rules at major CDNs sometimes block OAI-SearchBot even when site owners intend to allow it. Audit managed rule lists quarterly and confirm OAI-SearchBot is whitelisted.
• Hardcoding UA suffix versions. OpenAI revised crawler suffixes during 2025. Match on the substring OAI-SearchBot and ChatGPT-User, not on a specific version like OAI-SearchBot/1.2.
• Assuming referer attribution will work. In-Atlas summaries frequently do not produce clicks; relying on Referer: chatgpt.com for measurement undercounts true visibility (SALT.agency, 2025).

FAQ

Q: How is Atlas different from OAI-SearchBot?

Atlas is the user-facing Chromium browser; OAI-SearchBot is the backend crawler that indexes pages for ChatGPT and Atlas citations. Atlas presents an unmodified Chrome user-agent and is treated as human traffic, while OAI-SearchBot self-identifies with a +https://openai.com/searchbot suffix and obeys robots.txt (OpenAI Crawler Docs, 2026). Blocking one does not block the other.

Q: Should I rate-limit Atlas?

Generally no. Atlas operates inside a real user's session and looks indistinguishable from Chrome on macOS at the network layer (HUMAN Security, 2025). Rate-limiting Chrome 141 on macOS will hit legitimate human users alongside Atlas. If you must throttle, use behavioral signals (missing fingerprinting JS, suspiciously linear navigation, absent mouse events) rather than UA filtering, and apply throttles only on protected endpoints.

Q: How do I verify a real Atlas request?

You cannot reliably verify Atlas at the request layer because Atlas does not advertise itself and does not currently sign its traffic (HUMAN Security, 2025). The closest practical signal is a heuristic combination of UA match (Chrome/141 ... Safari/537.36), reasonable session length, ARIA-tree access patterns, and absence of automation flags. For verifiable AI traffic, focus on the upstream OAI-SearchBot and ChatGPT-User crawlers, both of which expose verifiable user-agent suffixes and IP ranges.

Q: Does Atlas honor my robots.txt?

Atlas itself, as a browser, does not consult robots.txt — browsers never have. The crawlers that feed Atlas citations do: OAI-SearchBot honors disallow rules for search inclusion, and GPTBot honors them for training (OpenAI Crawler Docs, 2026). OpenAI's December 9, 2025 documentation update clarified that user-initiated ChatGPT-User fetches may bypass robots.txt because they are treated as proxied user actions rather than crawls (ppc.land, 2025).

Q: Will blocking GPTBot remove me from Atlas answers?

No. GPTBot is the training-data crawler; blocking it prevents your content from entering future training corpora but does not affect real-time citations in Atlas (OpenAI Crawler Docs, 2026). Real-time citations are sourced via OAI-SearchBot and ChatGPT-User. Block GPTBot if you want training opt-out; allow OAI-SearchBot if you want Atlas and ChatGPT search visibility.

Q: How do I measure Atlas-driven traffic?

Reliable measurement requires three signals in combination: (1) Referer headers from chatgpt.com and OpenAI-owned domains, (2) log-level segmentation of the four user-agent channels described above, and (3) UTM-tagged outbound links wherever you control them. Note that in-Atlas summaries may answer the user without producing a click, so absolute Atlas-driven sessions will undercount true citation impact (SALT.agency, 2025).

Q: Does Atlas index my page on its own, or does it rely on OAI-SearchBot?

Atlas does not maintain an independent index. The citations rendered in Atlas's sidebar and agent answers are sourced from OpenAI's search infrastructure, which is fed by OAI-SearchBot. If OAI-SearchBot cannot reach or parse your page, Atlas will not cite it even if a logged-in Atlas user opens the URL directly. Treat OAI-SearchBot allow rules as the gating control for Atlas citation eligibility (OpenAI Crawler Docs, 2026).