Can ChatGPT Sites Replace a Web Designer?

Sites turns the request into a local web project, builds a deployable version, then publishes that saved version through OpenAI-managed hosting. OpenAI treats saving and deploying as separate stages, so a version can be reviewed before it becomes the production site.

Roadnote began with OpenAI's Sites starter and shadcn. Codex then changed the source, ran the app, built it, and passed the output to Sites hosting.

Roadnote arrives from Cloudflare with Next-style /_next/ assets, React Server Component data, and Vinext markers. Calling it a Next.js app is fair. Production behaves like a Cloudflare Worker.

Sites can also add D1 records, R2 files, identity, Sign in with ChatGPT, secrets, access controls, analytics, and custom domains. Roadnote needed none of them.

The first draft is unusually complete for a thirty-minute build. Its strongest work sits in product flow and presentation. The weak spots appear where a polished interface makes a promise about technology that still depends on the visitor's browser.

No. Roadnote is a working one-screen prototype, not a full business website or finished product. It has no navigation, supporting pages, Privacy Policy, Terms of Service, support route, visible company identity, or legal owner. It also needs correction, export, recovery, installability, security review, measurement, and edge cases.

To be fair, my one-sentence prompt did not ask for a full company website or a legal section. Sites met the narrow brief. The danger is that a narrow result can look finished enough to hide the scope nobody requested.

The missing Privacy Policy is more than a footer detail. Roadnote requests microphone access, relies on browser speech recognition that may use a remote service, writes notes into localStorage, and receives a Cloudflare bot cookie. A policy should explain each fact in language a driver can understand.

Terms should cover availability, acceptable use, the limits of speech recognition, responsibility for saved notes, and the obvious warning against handling a phone while driving. Navigation should connect those policies to how-it-works, browser support, help, and contact pages.

That does not cancel the achievement. “It works” and “it is ready for customers” are different sentences.

Roadnote's response set Cloudflare's __cf_bm bot cookie. Cloudflare says it is separate from the app and not tied to an app user ID. A broad “no cloud” impression still misses that network behavior.

The response lacked the security headers in the list above. CSP restricts resource loading. HSTS keeps future requests on HTTPS. Permissions Policy makes microphone access explicit. Missing them does not prove an exploit. It shows that hardening was not finished.

OpenAI ties Sites to local source and a Git commit, better than a sealed builder. Moving a product still means moving hosting, identity, data, secrets, analytics, DNS, and browser notes.

A ChatGPT Site can be indexed because the platform can return readable HTML with normal titles, headings, and descriptions. Ranking still depends on useful content, links, authority, technical cleanup, and whether the site answers a searcher's question better than the pages already competing for it.

Roadnote starts better than many JavaScript demos. Its headline and supporting copy appear in the first HTML response. Google recommends server-side rendering or pre-rendering because crawlers and visitors get useful content without waiting for JavaScript to assemble it.

A one-page site can be indexed without a sitemap, and Google can choose a canonical. Both matter more after adding routes, a custom domain, or duplicate addresses.

Purpose is the bigger problem. “Roadnote” has no search history. The page does not answer questions about voice notes, driving safety, browser support, or privacy. There is little to link to and nowhere to crawl next.

Ninety-one words can explain one button. Search needs compatibility, privacy, driving safety, dictation comparisons, and support answers based on real problems. Each page needs a reason beyond repeating a keyword.

AI code receives no automatic penalty. Google focuses on useful, original, first-hand material. Thin stays thin regardless of who typed it.

The social card improves shared links, not rankings. Until Roadnote moves, mentions build history around chatgpt.site instead of a domain the product owns.

Google can use og:image for a preview. It also says standard img elements aid image discovery, while CSS backgrounds are not indexed as page images. Roadnote offers no body image.

Truthful SoftwareApplication markup could describe the app, support, and price. It must match visible facts. Schema should not be richer than the page.

Sites traffic analytics cannot tell whether a note was saved. Search Console, indexing checks, queries, and a privacy-conscious completion event answer different questions.

Which ChatGPT Sites are already appearing in search results?

On August 29, 2026, an unsigned Google search from Tucson for site:chatgpt.site returned ten URLs on its first page and showed more result pages. The first three are listed below without direct links. This is a snapshot, not a complete index count. Google can change the order and results by date, location, and search history.

These three examples show that Google knows about some chatgpt.site URLs and can return them when the search is restricted to that host. That is not the same thing as being discoverable in an ordinary search.

I searched Google for the visible title and handle of all three results: “Actually Worth It? · nfl77,” “Baro Tool · comanee1980,” and “Dream Pattern Lab · akynori.” None of the three ChatGPT Sites appeared on the first page. A URL can therefore show up under site:chatgpt.site and still be difficult to find when somebody searches for its own name. It certainly does not prove that the page ranks for valuable unbranded searches, brings leads, or will keep its position. Indexing is admission to the race, not a result.

Even the first three show the range. One is a buying guide, one is a large utility collection, and one is a niche reference publication. Baro Tool has navigation and a privacy page. More prompting can produce more pages. Somebody still has to decide which pages are true, useful, connected, and worth maintaining.

Let's look at a second ChatGPT Site

I am not affiliated with The Oompa Meisters. This is a dated review of a public site on August 29, 2026, not a judgment about the band or its work.

My estimate is that this probably took someone four to eight hours of prompting, revisions, content gathering, domain setup, and cleanup. I cannot verify their time. I can see the larger scope. This is much closer to a real business website than my 30-minute Roadnote build, and it still left a practical to-do list.

The site has a clear offer, real band photography, a service area, testimonials, phone and email links, public events, service pages, merchandise, several booking buttons, and a private musician portal. Its custom domain is live, and the chatgpt.site pages point canonical tags, robots.txt, and the sitemap toward that owned domain.

I checked the links instead of trusting the score. Eleven public routes returned distinct HTML pages, and the proclamation PDF returned successfully. The main navigation is not fake. The interaction gap sits one level deeper.

The Upcoming Shows cards contain dates, locations, and times but no links for tickets, venue details, directions, or adding an event to a calendar. On the day of this audit, the August 20 and August 22 events were still presented as upcoming. The media page's “Open the YouTube channel” link opens one video rather than the named channel. These elements display information without finishing the visitor's likely task, which makes parts of the site feel like they do nothing.

The booking path does not work for the public. Every booking button I checked points to the same private Google Form. Following that exact link returned an HTTP 401 response and a Google sign-in page. A visitor cannot ask for pricing or check a date through the site's main conversion path.

Privacy and Terms routes returned 404 even though Google Analytics loads and the musician portal asks for an email address and PIN. More seriously, the login screen tells musicians to use the first six digits of their phone number as that PIN. I did not submit the private form, but that published credential rule is predictable and should be replaced with proper authentication before launch.

The Lighthouse scores still deserve credit. The pages have descriptive titles, useful metadata, canonical tags, a crawler file, a sitemap, skip navigation, sensible landmarks, and good alt text. I found no JSON-LD on either the canonical home page or events page. The site has no structured entity for the band and no Event objects for its public performances.

JSON-LD is not required for ordinary indexing. It is the clearest way to tell search engines that this is a MusicGroup or Organization and to describe each show with a name, date, status, venue, address, performer, image, URL, and ticket offer when one exists. Internal pages would also benefit from BreadcrumbList. The visible page must support every field. Markup cannot manufacture missing venue or ticket details.

Adding JSON-LD to the current schedule page would not finish the job. Google's event guidance says each event needs a unique leaf URL with its structured data on that page. The Oompa site has one list of static cards. We would give each public show a useful page, connect it to tickets, directions, and the venue, then add accurate Event markup there.

That makes the 100 SEO score narrower than it looks. It means the tested page passed Lighthouse's SEO checks. It does not mean the site has the structured data needed for event eligibility, a complete local search plan, working conversion paths, or fresh listings.

This is exactly where Internet Crafters earns its keep. A production review asks whether a customer can complete the job, whether old information is retired, whether private access is actually private, whether analytics has a policy, and whether the owner knows what to update next month. Generating twelve polished routes is impressive. Taking responsibility for the whole system is the website work.