The expectation is usually this: you write to ChatGPT “make me a site for a coffee shop”, and in a couple of minutes the site is ready. Reality is a little more complicated. In a couple of minutes a lot really will be ready: the structure, the texts, the page code, sometimes even a preview right in the chat. But that is a draft. Between it and a site that opens at an address, accepts requests and looks normal on a phone, there are several more steps, and ChatGPT does not do them by itself.
I will sort out what ChatGPT can do when a site is created, where a person’s work begins, and what the path from an answer in the chat to a draft page is made of.
What ChatGPT does when a site is created
Structure and texts
This is worth starting with, even if you are not going to write the code. ChatGPT is good at laying a site out into pages and blocks: what is on the home page, which sections are needed, in what order the landing blocks go, what to write on the first screen.
It writes texts quickly too, but the quality depends on what you gave it. If the request is only “a coffee shop in the city center”, the texts will come out general. If you give the menu, the particular features, the address, the opening hours, who your guests are and how you differ, the text becomes more concrete. ChatGPT takes facts from the request, and when they are not there it may invent them, so prices, deadlines and promises in the text have to be checked.
A good draft starts from questions, not from an answer: who the audience is, what the goal of the page is, how the business differs from its neighbors, how to get in touch. ChatGPT can ask such questions if you ask it to. But understanding which answers really matter for the page, and which only sound convincing, is already the work of a person who has seen many such projects.
HTML, CSS and JavaScript code
ChatGPT writes the page code: HTML markup, CSS styles, simple scripts in JavaScript. For a one-page site it can return everything in one file that will open in a browser.
The code is usually working for simple tasks: blocks, a menu, buttons, a gallery, an accordion of questions. It is harder with what needs a server: sending a form, a personal account, payment. Here ChatGPT will write the code, but without a configured server and services that code will not send anything.
A separate point is images. ChatGPT does not know what your room, product or team looks like. In the code it will put placeholders or links to pictures you do not have. Photos and graphics have to be prepared separately, brought to a reasonable size and put into the code. If that is not done, the page will be either empty or full of other people’s pictures.
One more detail: code from the chat lives in the chat. If you ask for edits several times, it is easy to get lost about which version is the last, and to lose successful changes. Without the habit of keeping versions that developers have, this happens almost always.
Preview in the chat
OpenAI’s help says that for supported code blocks you can switch from the code to a preview right in ChatGPT. Among the supported ones are HTML pages and React components. The set of available actions depends on the plan, the device, the workspace settings and the model, so the preview button may not be there.
A preview is convenient for seeing quickly how the page looks and asking for a fix. But it is a view inside ChatGPT: the site does not appear on the internet because of it, and a check on a phone and in different browsers is still needed separately.
From an answer in the chat to a page on the internet
Hosting and the domain
For the page to open for other people, the files have to be placed on hosting and a domain attached. ChatGPT will explain how to do it and suggest options, but it places nothing itself: choosing hosting, registering a domain, uploading the files and setting the address has to be done by hand.
After placement it is worth checking that the site opens at a protected address with https, that the address with www and without it leads to the same page, and that all the pictures and styles loaded. A frequent mistake is paths to files that worked on the computer and broke on the hosting.
Forms
A request form in code from ChatGPT usually looks finished: fields, a button, a “Thank you” message. But for the request to arrive somewhere, a handler is needed: a server, a form service, or an integration with email, Telegram or a CRM. Without that, the form neatly sends data into nowhere. This is one of the most frequent traps of homemade sites: everything looks working until it turns out that the requests were not arriving.
This is checked only with test requests from a phone and from a computer: both must arrive where they should, with all the fields.
If the form collects a name, a phone or an email, consent to the processing of personal data is also needed. That is no longer a question of code.
A check on a phone and in browsers
ChatGPT writes adaptive layout if you ask, but it cannot check it on real devices. In practice, on a phone, menus, wide tables, long words in headings, and buttons that end up too close to each other break most often. It has to be checked on a real phone and in several browsers.
What a draft is actually made of
The path from the request to a draft looks short only from the outside. In fact it has several stages: formulate the task and the action the visitor should take, assemble the structure, give the facts — services, prices, address, differences — get the texts by block, then the code, and look at the result. At each stage a person’s decision is needed: which blocks are extra, which facts must not be left to the model, what exactly to fix when the page “is not liked”.
People usually go wrong in the same places. They do not give facts, and the neural network invents prices and promises. They ask to “make it prettier” instead of a concrete edit. They take the preview for a finished site. They consider the form working although not a single test request has arrived yet. So what comes out is a draft, not a site, and then it is decided what to do with it.
Where the limits begin
| What is needed | ChatGPT | A person |
|---|---|---|
| Structure and texts | Suggests variants quickly | Gives the facts, checks and chooses |
| Page code | Writes HTML, CSS, JavaScript | Checks, edits, keeps track of changes |
| Preview | Shows it in the chat, if it is available | Checks it on real devices |
| Hosting and domain | Explains the order | Buys, sets up, places |
| Forms and requests | Writes the form code | Connects a handler and checks delivery |
| Personal data | Can draft a consent text | Is responsible for how it is arranged and stored |
| Changes later | Rewrites the code on request | Watches that the edit does not break the rest |
The main limit is not in the code, but in the fact that a site is not one file. It is hosting, a domain, forms, analytics, updates, backups. ChatGPT helps in each part, but a person ties them together and answers for the result.
There is also what ChatGPT will not remember if you do not ask: the Title and Description of the page for search, connecting analytics, a privacy policy, a clear error page if the address is typed wrong. Separately these are small things, but without them the page stays a draft even if it looks finished.
The second limit is changes. The first version of a page is assembled quickly. A month later the prices have to be changed, a block added, the form fixed. If the code was written entirely by a neural network and you do not understand it, each edit turns into a new request and a new risk of breaking something. The more edits, the harder it is to keep the code in order.
When to move a draft to Tilda or into development
A draft from ChatGPT is valuable even if the code is not used in the end: the structure and the texts are half the work on a site.
Leaving the page on code from ChatGPT is also possible. It is reasonable if this is a temporary page for one event, if you are ready to deal with the code and the hosting yourself, or if the page does not collect requests and is not updated. In all other cases it is worth thinking in advance who will change it and how in six months, when the prices change and the author of the code no longer remembers what is where.
Moving it to Tilda is worth it if the site is simple, it will have to be updated regularly by hand, and forms and requests matter. On Tilda I assemble pages from standard blocks and Zero Block, and a draft from ChatGPT, if there is one, helps me understand what matters to the client. I still build the structure and the texts for the task myself. I do not transfer finished HTML code from a neural network onto Tilda: the page is then hard to maintain.
This approach has a practical plus: a page on Tilda can be changed without code. Prices, texts, photos and the order of blocks are edited in the editor, and forms are connected to email, Telegram or a CRM with the standard settings.
It is worth handing it to development if non-standard logic is needed: a calculator, a personal account, integrations, non-standard forms. Then a draft from ChatGPT becomes a visual brief, and it is easier for a developer to understand what you want. For such a task, custom development fits, and if one selling page is needed — a landing page.
If you want to discuss what to do with your draft, take the quiz on mara-craft.com. I will look and suggest which path fits.