ChatGPT Sites Generates Over 5 Million Sites in Three Months
OpenAI announced that approximately three months after the launch of ChatGPT Sites, users have created over 5 million sites. The platform has introduced collaborative editing, targeted private sharing, reduced deployment time by about half, the ability to check site databases through conversations, and the option to bind custom domain names.
This feature initially entered commercial and enterprise workspaces in preview form on June 2, 2026, allowing Codex to turn prompts into hostable lightweight websites or applications, hosted on Cloudflare, with ChatGPT login and file storage. On July 9, it expanded to paid tiers and allowed public publishing. Starting August 20, site owners can add members within the same workspace as editors, allowing them to modify, save, and push new versions to the same URL after the owner's initial publication; ownership, keys, domain names, and audience settings remain with the owner. Private sharing allows external visitors to be invited via email, requiring them to log in with the invited account without needing to join the workspace or make the site public. Custom domain names were not available at the launch of enterprise workspaces.
Official examples include dashboards, portfolios, onboarding pages, Q&A games, local maps, and itineraries. Demonstration sites feature pixel exploration games, 3D coloring in browsers, summaries of earthquakes and space weather, sound-enabled drama calculators, and Japanese word chain games. Documentation states that Sites is in public testing, targeting Plus, Pro, Business, Enterprise, and Edu users. Codex has previously reported over 5 million weekly active users, of which about 20% are non-coding knowledge workers, growing faster than traditional engineers.
Buyers are teams and individual creators looking to write fewer deployment scripts; sellers are OpenAI, which bundles generation, hosting, and permissions into a single subscription. The driving force is the permission gap following the 5 million sites, not a new model. Funding comes from paid seats covering hosting and agent usage. Beneficiaries are teams that can deliver internal tools using their own domain names; those under pressure include low-code platforms that charge hosting fees per project and traditional website building processes that still require moving applications out of conversations to go live.
Source: Public Information
ABAB AI Insight
5 million refers to the creation count, not unique visitors. OpenAI aims to prove that conversations can directly turn into shareable runtimes, rather than just remaining as chat logs. After three months, the additions of multi-user editing, private lists, and domains are crucial—without these three, sites are merely demonstrations; with them, they can integrate into client links and team meetings. The halved deployment time reduces trial and error costs from 'opening another repository' to 'issuing another prompt.'
The capital path is subscription covering hosting. Cloudflare provides the network, Codex handles generation and backend Git, and ChatGPT manages identity. The initial restriction on custom domain names for enterprise accounts separates branded sites from internal sandboxes. The faster growth of knowledge workers compared to engineers indicates that Sites is targeting the layer where spreadsheets and slides transform into links, rather than competing with production applications on Vercel.
In comparison to Bubble, Glide, and early Google Sites: low-code platforms charge based on component markets, while conversational site building relies on seats and tokens. The industry phase has shifted from 'being able to generate pages' to 'being usable as internal systems.' Databases can be viewed in conversations, and editors can also see them, with the permission model beginning to resemble a lightweight backend rather than static pages.
The structural change involves technology replacing intermediate steps and a shift in pricing power. Replacement occurs in the steps from design drafts to launch; pricing power shifts from domain registrars and website packages to conversation subscriptions that already have login states. The mechanism is: when publication, permissions, and data are all contained within the same account system, the cost of switching platforms is no longer code, but rather the list and domain resolution.
ABAB News · Cognitive Laws
Generating pages is not enough; managing permissions counts as a product.
Creation volume proves attempts; domain names prove delivery.
When conversations turn into runtimes, the first intermediary to be removed is the publication step.