Can you build a website with ChatGPT? We tried

“Just ask ChatGPT to build your site” has been the most repeated advice of the last two years. You would expect a website-builder company to say “no, that does not work”. But it does work — the question is what works, and exactly where it stops. Let us be honest about it.

Why everyone is asking

There is a reason this question is everywhere: the “I built a website with ChatGPT in five minutes” videos. Those videos are not lying — they really do build something. What they do not show is that the video ends exactly where the work begins. No domain, placeholder images, a contact form that goes nowhere, and a site that is not published anywhere at all.

So let us take the question seriously and follow it all the way through. What follows is not one bad attempt; it is what falls out of how the tool is shaped.

The short answer

Yes, ChatGPT will give you a working web page. One file, it opens, it can even look good. No, that is not a business website — because a website is not a file. It is a living thing: it has to sit at an address, be updated, be findable, and still work in six months. A chat window does none of that.

What it genuinely does well

Credit where it is due — this list is serious:

• Writes the HTML and CSS for a single landing page in minutes
• Writes your copy, which removes the blank-page problem entirely
• Teaches you how a component works — it is excellent for learning
• Finds and explains the bug in code you already have
• Produces a quick prototype so you can show a client “something like this”

These are not small things. A meaningful slice of a designer’s week is on that list. The problem is not what ChatGPT can do — it starts afterwards.

Where it hits the wall

More than one file

The first page comes out well. Ask for a second and the navigation is slightly different. On the third the button colour drifts. By the fifth you are saying “make the header the same on all of them” and the thing has become a patchwork. The reason is simple: a chat is a conversation, not a project. Every request starts in a slightly different world.

Images

The code says “an image goes here” and no image arrives. Producing images is a separate job, sizing them is another, optimising them another again. This is also where the number-one cause of slow sites is born: a four-megabyte hero photograph.

Publishing

You have a file. Now what? You need a domain, hosting, DNS records and a certificate. ChatGPT will explain all of it — but it will not do any of it. The “five minute website” story usually ends right here, in hour three.

The second-day problem

This is the big one. Three months later your phone number changes. You find the old conversation and say “change this”; the model does not remember that site, and you are no longer certain which file is the one that is live. It produces a new version, you try to upload it, something breaks. Building a website takes a day; owning one takes years — and that part does not exist in a chat window.

The technical layer you cannot see

Ask for them and it will write them: canonical tags, Open Graph data, sitemap.xml, robots directives, schema.org, mobile breakpoints. But you have to ask — which means you have to already know what to ask for. If you do not, those things quietly stay empty, your site does not show up in search, and nobody tells you anything is missing.

That is a general property of these tools: they answer the question you asked, they do not remind you of the one you did not. A template makes the unforgettable things unforgettable; a chat window trusts your knowledge. For a designer who knows the territory that is fine. For a hairdresser building their own site it is the difference between existing in search and not.

Accessibility and mobile

The output looks fine on a desktop. On a phone there is a good chance of an overflowing button, a link too small to tap, or a form you cannot get through with a keyboard — because the model does not test its output, it only produces it. Think of a designer who writes without ever seeing the result: capable, but blind.

Doing the same job twice

You built a site for one client; next week you build a similar one for another. You reopen the conversation, type the same requests, and get a different result — because every generation is a fresh roll of the dice. For anyone trying to build a repeatable workflow, that is a quiet cost that eats the speed gain.

ChatGPT makes web pages; it does not make websites. The difference is not one word — it is the domain, the hosting, the maintenance and the six months afterwards.

What happened when we tried it

Take a concrete example: “build a site for a veterinary clinic in a small town”. The first output is better than you expect — a decent hero, service cards, a contact form, all in one file, all reasonable. It is entirely understandable that people watch this happen and conclude that building a site is easy now.

Then the real work starts. “Add a team page”: a second file arrives, but its navigation differs slightly from the first. “Add the logo”: there is no logo, only a placeholder. “Make the booking form work”: the form is there but submitting it goes nowhere, because there is no backend. “Publish it”: it explains the steps, in order, and does not perform any of them.

So the first eighty per cent genuinely arrives fast. The remaining twenty — consistency, images, a form that works, publishing, the domain — takes as long as the whole job. Software people have called this the last twenty per cent for decades, and websites do not break the rule.

“But I know how to code”

If you are a designer who codes, most of the obstacles above are not obstacles. You take the file, drop it into your project, fill the gaps and publish. ChatGPT is a genuine accelerator for you, and this article is not telling you to stop.

But do the arithmetic: how many hours does it take you to make that file usable, publish it, and update it six months later? That cost does not happen once — it repeats on every project. For a designer working at volume the real question is not “can I do this” but “do I want to do this again for every client?”

Who it is genuinely enough for

There are cases where it really is enough, and there is no point pretending otherwise:

• A single-page personal site that will never be updated
• A quick prototype to test an idea
• Learning — understanding how the code works
• As an accelerator for someone who already codes and will handle publishing themselves

If you are not on that list — if you are delivering work to a client, or your business site will need updating — ChatGPT is a starting tool rather than a destination.

The real use for a designer

Do not read this far and conclude “so I should not use it”. That is the wrong conclusion. Using ChatGPT as a website builder is weak; using it as an assistant is strong. The difference is what you expect from it.

• First drafts of client copy — the blank-page problem, solved
• Learning a CSS property you do not know, in five minutes
• Pasting an error message and understanding what happened
• Softening a difficult email to a client
• Showing quickly what an idea might look like

There are several hours a week on that list. Forcing it into the “site builder” role is what leads to disappointment — like driving a screw with a hammer. The problem is not the hammer.

The real comparison

The comparison is not really “ChatGPT versus a site builder”. It is between a general-purpose chat and a tool that does one job end to end. What a builder adds is not a cleverer model; it is everything missing around the model — a template designed by human designers, hosting, domain connection, version history, and technical SEO arriving by default rather than on request.

An ironic footnote: builders run language models too. So the difference is not the intelligence of the model. It is whether that intelligence is working on a blank page or on a proven skeleton, and whether the things a site needs in order to actually live are there around it.

What if your client tries it?

Sooner or later a client will say “I will just get ChatGPT to do it”. Do not panic, and do not try to talk them out of it — the side arguing against it looks defensive.

Usually one of two things happens. Either they spend a week on it, fail to get it published, and come back — this time understanding what you were worth. Or they succeed, in which case they were never going to be your client anyway, and you have lost the lowest-budget, most time-consuming job on your list. Both outcomes are in your favour.

In short: you can make a web page with ChatGPT, and a good one. A published website that can be found and can still be updated in six months is a different thing. The first is an afternoon’s work; the second needs a tool.