AI · Models & tools
ChatGPT Sites: The Builder Europe Doesn't Get
Since July 2026 ChatGPT builds, hosts and publishes websites straight from the chat – just not in the EEA, Switzerland or the UK. What Sites can do technically, what its contract forbids, and why the blocked region is the more interesting story. As of July 2026.
By Boaz Lichtenstein Prefer us on Google
There is a rare kind of product launch: the one everybody writes about and nobody here can open. ChatGPT Sites, generally available since 9 July 2026, is one of them. You describe a website or a small web app in the chat, ChatGPT builds it, hosts it and hands you an address – draft, deployment and operation in a single window. Just not in the European Economic Area, not in Switzerland and not in the United Kingdom. Here is what the tool does, what its contract forbids, and why the blocked region is the more interesting story. As of July 2026.
Key takeaways
- Sites is unavailable in the EEA, Switzerland and the UK – with no announced timeline and no explanation in the release notes.
- It grew out of “Codex Sites” (June, workspaces only) and has been part of ChatGPT Work since 9 July on all paid plans except Free and Go.
- Technically it runs on Cloudflare Workers with a D1 database and R2 file storage; every deployment URL is production, there is no staging.
- Selling is contractually forbidden: no payments, no card data, no financial or crypto transactions, no health information.
- Published pages live at
<name>.openai.chatgpt.site– an address that looks like a sender but is not one.
What Sites is: from chat to hosted website
The origin lies in June 2026, when OpenAI launched the feature as a
workspace beta inside Codex, invoked with @Sites. On 9 July it became
ChatGPT Sites: available on the paid plans except Free and Go,
created in ChatGPT Work in the browser or in Work or Codex in the
desktop app.
The flow is deliberately short. You describe what should exist, attach files, data or links, and get a preview visible at first only to you and your workspace admins. Changes happen in conversation. On publishing you pick the reach – just you, selected people, the whole workspace, or the open internet – and receive an address.
One detail deserves particular attention: every deployment URL is a production URL. There is no test environment alongside. If you want to review something before it goes live, you have to explicitly ask ChatGPT to save a version without rolling it out. Anyone who does not know that publishes their first draft.
Sites is clearly positioned for the short path from idea to address: dashboards, project trackers, launch calendars, internal portals, reports, prototypes, small calculators and reference pages. In other words, the things that currently sit around as a spreadsheet with six tabs and that nobody would set up a project for. That is an honest niche – and it explains why the feature set underneath is cut the way it is.
The restriction: EEA, Switzerland, UK
OpenAI keeps it brief in its help centre: at launch, Sites is not available in the European Economic Area, Switzerland and the United Kingdom. No schedule, no explanation, and July’s release notes never return to it. For users in those regions it simply means the feature does not appear, even though ChatGPT Work itself runs normally.
OpenAI does not say why – and anything further would be speculation. As context rather than explanation: Sites moves OpenAI from a provider that generates answers to one that hosts and publicly serves third-party content. That touches a different set of obligations – responsibility for hosted material, complaint routes, transparency requirements – than a chat window triggers. A documented limitation fits the picture: Sites supports neither data nor inference residency, precisely the assurances OpenAI otherwise offers European enterprise customers. The connection is not proven; it is the obvious reading.
What’s inside – and where the limits are
Under the surface, Sites is built more seriously than the chat entry point suggests. It runs on Cloudflare Workers, gets a relational database in D1 and an object store for files in R2; sign-in works via workspace identity or “Sign in with ChatGPT”. That allows for small applications with real data, not just brochure pages.
| Works | Doesn’t work |
|---|---|
| Server-side logic on Cloudflare Workers | Node servers, Postgres, external databases |
| D1 for records, R2 for files and uploads | Background jobs, long-running processes, WebSockets |
| Sign-in via workspace or “Sign in with ChatGPT” | Data or inference residency |
| Your own domain via DNS records | Your own domain in Enterprise workspaces (at launch) |
| Saving versions instead of deploying immediately | Restoring a deleted Site |
On top of that come plan-specific beta limits on storage, traffic and number of Sites: once reached, nothing new can be created or made public until usage falls. And deletion is permanent, with no restore.
Selling forbidden
The hardest limit is not a technical one. OpenAI explicitly prohibits Sites from handling payments, processing card data, and enabling financial and crypto transactions; health information and offerings aimed at children under 13 are barred as well.
That answers the most obvious question: this will not be a shop. Catalogues, product galleries, landing pages, internal portals, dashboards, project trackers, calculators – yes. The checkout belongs elsewhere. Anyone already thinking about how commercial content becomes discoverable to AI systems will find the relevant perspective in LLM optimisation for stores.
The trust problem in the address
One point gets too little weight in the coverage. Published Sites live
on a shared hosting domain, and in OpenAI’s own documentation it looks
like this: goblin-tales.openai.chatgpt.site. The pattern is
<name>.openai.chatgpt.site – the “openai” part belongs to the
platform, not to the sender. It says nothing about who built the page.
Most people, however, read an address left to right as a statement of origin. A page on a domain containing “openai” looks more official than it is. For brands that means it is worth watching your own name in this environment. For everyone else, the old rule applies in a new setting: the trustworthiness of a page comes not from what stands in the address but from a legal notice, a verifiable identity and a sender you can confirm independently. How quickly convincing fakes are produced today is covered in spotting deepfakes.
What to use instead
For the blocked regions the unromantic answer is: the same tools as before, with a grown AI share. Three categories cover practically everything Sites is meant for:
- Classic site builders with AI assistance – fastest for marketing pages, with European hosting options and no ban on selling.
- AI builders that emit code, where the result is a project you host yourself. More effort, no lock-in risk.
- Static hosting plus a coding agent – the route this site itself takes. Most control, least convenience.
The common denominator: all three hand you the files. That is precisely the point on which Sites makes no promise today.
The bottom line
ChatGPT Sites is technically more serious than its chat entry point suggests – database, file storage, sign-in and custom domains are not a toy. It is also a product with three hard edges: it may not sell anything, it guarantees no exit path, and across the blocked regions it does not exist at all.
That makes the recommendation an unexcited one: note it, do not plan around it. Anyone wanting to understand the trend should look at the actual movement – the large AI providers are ceasing to supply tools only and beginning to be the platform underneath. What that means for working with them is covered in ChatGPT Work.