ChatGPT Sites, Explained: Who Can See a Site and How to Get Feedback on One (2026)
By Marcin, founder of Markloop · Last verified September 2026 against OpenAI's Sites docs and help center
Short answer: ChatGPT Sites (it started in June as Codex Sites) builds a website or small web app from a chat and hosts it at a URL you can share with your workspace, with people you invite, or with anyone on the internet. It's a public beta on Plus, Pro, Business, Enterprise and Edu, not Free or Go; Plus and Pro users in the UK, EEA and Switzerland got it on July 20, 2026, and for workspace plans there, check OpenAI's help center for your country. Viewers can't comment on a Site, so for a proposal or report you need a response to, publish that page somewhere people can comment.
What ChatGPT Sites is (and what happened to Codex Sites)
You describe a site in a chat, ChatGPT builds it, and it gets hosted at a URL without a separate deploy step. OpenAI's Sites documentation covers landing pages and content sites, sites that save records or game scores, sites that take uploads, and internal tools that know which workspace user is signed in. Under the hood a Site can keep data in a database (up to 10 GB) and files in object storage, and it speaks HTTP, HTTPS and WebSockets.
OpenAI announced Sites inside Codex on June 2, 2026, as a preview for Business and Enterprise workspaces, alongside Codex Annotations. In July it came to ChatGPT as a public beta, and the docs moved from the Codex developer site to ChatGPT's own. Today it's one product: you can build a Site in a ChatGPT chat, or start from a local project in Codex, which links the project to the Site in .openai/hosting.json. People still search for "Codex Sites", and it's the same thing.
The basics (September 2026):
| Question | Answer |
|---|---|
| Plans | Public beta on ChatGPT Plus, Pro, Business, Enterprise and Edu |
| Price | Included in the plan, with plan-specific usage limits shown in the product |
| Regions | UK, EEA and Switzerland: Plus and Pro since July 20, 2026; for Business, Enterprise and Edu there, check OpenAI's help center |
| Versions | Save a version to review it, then deploy it; the URL stays the same |
| Custom domain | ✓ you own the domain and add DNS records (OpenAI's docs still exclude Enterprise workspaces) |
| Analytics | ✓ unique visitors and page views (not for Enterprise workspace-owned Sites) |
| Viewer sign-in | Optional "Sign in with ChatGPT" on public Sites |
| Not allowed | Health records (PHI), payment-card data, sites aimed at children under 13, financial transactions |
Who can see a Site
The owner picks an audience, from narrow to wide:
- Owner and workspace admins - private, the default while you build.
- Selected people or groups in your workspace.
- Invited external viewers - people outside your workspace you invite by email. They don't become workspace members, and they have to sign in with the account you invited. OpenAI describes this as rolling out to Plus, Pro, Business and Enterprise.
- Anyone in the workspace.
- Anyone on the internet - a public URL. In Enterprise workspaces an admin has to turn public publishing on first.
You can also add workspace members as editors. They can change the Site and publish updates after your first publish, but they work through ChatGPT like you do, and they can't change the audience or restore older versions.
That's more reach than most agent tools give you. A Cursor shared canvas only opens for your Cursor team, and Claude's public links depend on your plan (Claude's sharing rules, explained).
What a Site can't do: let people respond
Every audience option above is about viewing. The Sites docs describe no way for a viewer to comment on the page, point at a paragraph, or answer a question. The feedback features OpenAI shipped are for the author:
- Codex Annotations (June 2026) let you select a region of a document, slide, image or Site and ask Codex to change just that part. It's for instructing the agent; your client can't use it.
- Editors change the Site through ChatGPT. That's co-authoring, and it only works for members of your workspace.
So the client reads the Site and replies by email: "the timeline in the second part feels tight, and can we drop the blog?" You paste that into ChatGPT, work out which part they meant, and deploy again.
ChatGPT Sites vs. the other ways to share agent output
| ChatGPT Sites | Claude Code artifacts | Cursor shared canvas | Markloop | |
|---|---|---|---|---|
| People outside your org can open it | ✓ public URL or email invite | ✓ public link, depending on plan | ✗ your Cursor team only | ✓ share link |
| Viewer needs an account | public: no · invited: yes | public link: no | yes, on your Cursor team | no - they type a name to comment |
| Viewers can comment on the page | ✗ | in-org only, not on public links | ✗ | ✓ anchored to the element |
| Feedback gets back to your agent | author's own Annotations | ✓ Claude Code reads in-org threads | Design Mode, author only | ✓ over MCP from ChatGPT, Codex, Claude, Cursor, Grok |
| Hosts a real app (database, uploads) | ✓ | ✗ one static page | ✗ | ✗ documents only |
If the Site is a working app - a tracker with saved records, a form that takes uploads, an internal tool - Sites is the right home, and Markloop doesn't host apps. If it's really a document, like a proposal, a report or a one-pager, what's missing is a way for the reader to answer on the page.
Getting comments on a proposal or report you made with Sites
This is for document-like pages only: a proposal, a monthly report, a one-pager for a client. You don't move the Site anywhere; you publish a copy of that one page to Markloop, where the client can comment on it. In the chat where you built it, ask:
Save the Q3 report page of this Site as one HTML file with inline CSS and publish it to my Markloop project as a new document.
Send the link with comments on. When Priya has commented, ask ChatGPT to pull the comments and publish the next version; the link she already has opens it. ChatGPT connects to Markloop with a normal sign-in to your Markloop account, no Developer mode. In Codex it's two commands:
codex mcp add markloop --url https://app.markloop.io/api/mcp codex mcp login markloop
No setup needed to try it: download the HTML file ChatGPT made and upload it to Markloop. Watch the 2-minute walkthrough.
Sending a proposal? Here's how to write a client proposal with AI and get comments on one link, step by step. Who can see your link and where your data lives is in the FAQ.
FAQ
Is ChatGPT Sites available in the UK and the EU?
Yes for Plus and Pro: OpenAI added users in the UK, the EEA and Switzerland on July 20, 2026. Free and Go plans don't include Sites in any country. For Business, Enterprise and Edu workspaces in those regions, check OpenAI's help center for your country.
What changes when I save a Site's page as one HTML file?
It becomes a static snapshot. Anything that reads from the Site's database, depends on who's signed in, or submits a form stops working in the copy. For a proposal or report that's fine: everyone comments on the same fixed version.
Does my client need a ChatGPT or Markloop account to comment?
No. They open the Markloop link in a browser and comment. The name they give is self-reported, not verified. You can give the link an expiry date when you create it, and revoke it at any time.
Does the client get an email when the next version is up?
No. You can get an email digest of new comments (turn it on in your Markloop profile settings), but nobody gets an email about a new version. The client's link opens the latest version when they load or refresh it, so tell them when v2 is ready.
Where does Markloop keep the page and the comments?
Markloop is hosted in Europe. It stores the document and the comments; no AI runs on Markloop's servers. The edits are made by ChatGPT or Codex, not by Markloop.
Made a proposal or report with ChatGPT Sites?
Publish that page to Markloop, send one link your client can comment on, and let ChatGPT pull the comments into the next version.
14-day free trial · No credit card required · Reviewers always free
Related: Claude artifacts: the Claude equivalent · AI agent artifact sharing compared: Claude, Cursor, Codex · How to share Claude artifacts outside your organization · How to get comments on a Cursor canvas