The interesting thing here is not that ChatGPT can build a website; plenty of tools do. It is that the artifact you hand to colleagues stops being a static file you export and email and becomes a permissioned, living surface you regenerate in place. A slide deck is a snapshot the moment you attach it. A Site is a URL that shows whatever you last published — and that difference is the one worth thinking through.
OpenAI did two things at once on 9 July. First, it collapsed Codex into a single ChatGPT desktop app for Mac and Windows, so Chat, Work, and Codex now live in one place — with Work, the agent that goes off and returns finished spreadsheets, slides, documents, and web apps, available on every plan including Free. Second, and more interesting, it opened Sites in public beta: ask ChatGPT (or tag @Sites) to build you an interactive website or lightweight app, preview and refine it in the chat, then publish it to a URL.
Sites is not a chat transcript with a share button. OpenAI's own examples are dashboards, project trackers, internal portals, launch calendars, and reports — working surfaces, not documents. You can spin one up from Work or from Codex, edit it later by opening the conversation and describing the change (copy, layout, data, styles, forms, interactive behavior), and republish to the same production URL. It is in beta on paid plans, rolling out from Pro and Enterprise first, and — a real constraint — is not available in the EEA, Switzerland, or the UK at launch.
The part that unlocks new use cases is mundane on paper: access control. A new Site is limited to its owner and workspace admins until someone changes that, and in Enterprise workspaces public publishing is off by default and has to be switched on by an admin. In other words, the natural first audience for a Site is your own workspace — people who are already authenticated through your SSO — not the open internet. Sharing lets them view the live surface; it does not let them edit it.
That is the difference between this and a public website builder. You can hand a colleague a link to an interactive, current view of something — a metrics dashboard, a diligence index, a project tracker — without emailing a file, standing up a hosted app, or exposing anything publicly. The awkward middle ground most teams live in today, where "internal but interactive" means a spreadsheet nobody trusts is current or a Retool app somebody has to build and maintain, is exactly what this collapses.
Calling it a PowerPoint killer overstates it — decks are not going anywhere for a linear story you present in a room. But it is aimed squarely at the deck's worst habit: the moment you export to PDF and attach it, the numbers are frozen and wrong by the next morning, and every recipient now holds a slightly different stale copy. A Site inverts that. There is one link, it shows what you last published, and updating it is a sentence to ChatGPT rather than a re-export, a re-upload, and a fresh round of "latest version attached."
For the recurring artifacts that eat real hours — the weekly management dashboard, the board pack, the investor or data-room index, the status portal — a surface that stays at one address and refreshes on request is simply a better shape than a file that has to be regenerated and redistributed each time. That is the genuine unlock the merge and Sites deliver together: Work produces the material, Sites gives it a permissioned home that does not go stale the instant you send it.
The low-risk way to test this is to take one artifact you already rebuild on a schedule and try to live it as a Site instead: a management dashboard or an internal project portal, shared inside the workspace, refreshed weekly by prompt. Measure whether "one current link" actually beats "a new file every Monday" for your audience — for many recurring internal reports it will, and the switching cost is close to zero.
Two cautions before it touches anything sensitive. First, be precise about what "dynamic" means today: you update a Site by asking ChatGPT to regenerate and republish it, which is fast but is not the same as a dashboard wired to a live data source that refreshes on its own — treat auto-binding to systems of record as a thing to verify, not assume. Second, a data room is confidential by definition, and building one here means feeding diligence material into a hosted model and trusting a workspace access boundary to hold; for regulated or deal-sensitive content, confirm where the data goes and who can reach the link before you put anything real in it — and note the EU/UK availability gap while you are at it.
The headline is a super-app; the substance is a change in what a deliverable is. When your team's output can be a permissioned, updatable surface behind your existing login instead of a file that is stale the moment it is sent, the reflex to "export it and attach it" starts to look like a habit rather than a requirement. Not every document should become a live link — but the ones you keep rebuilding and re-emailing are telling you exactly which ones should.
The interesting move is turning a report you rebuild every week into one link that stays current — and being deliberate about the data you put behind it. If your team is trying to move an AI use case from demo to deployment, METECH helps scope, build, and validate the first working system in 2-3 weeks.