AI will build you a website in twenty minutes. Clean structure, decent layout, usable copy. That is not a promise anymore, it works. Anyone claiming otherwise has not tried it in over a year.
And it does more than build. Wired up properly, it also puts the site online, points your address at it and ships changes. That is not science fiction either.
Then it says: done.
That is the moment this article is about. “Done” initially says that the task has been worked on. Whether the result works needs verification. With browser access and suitable tools, an AI can inspect the delivered website and test forms. If it only reviews its own source code, it misses what actually reaches your customer.
A more expensive model does not replace that access. We have had pages reviewed by an AI several times over, thoroughly, with instructions to find faults. Afterwards a simple program ran across them that does nothing except look at the finished page the way a visitor receives it. It found faults on the first pass that every review round before it had waved through. Not because it is smarter. Because it stands somewhere else.
That is the whole difference. Not who builds. Who looks afterwards, and from where.
Spend enough time around this and you notice the problems keep taking the same shape. Four of them.
Faults That Never Announce Themselves#
We are used to computers turning something red when it breaks. With a website that is mostly not the case.
The most common example: someone fills in your contact form. The screen says thank you, we will be in touch. The message never reaches you. Nobody finds out. Not you, because nothing arrives. Not the person who wrote it, because they saw the confirmation. They wait two days and go to the next one.
The tricky part is not the fault. The tricky part is that a form losing enquiries looks exactly like a form nobody is using. A quiet week and a broken line feel identical. And nobody calls to tell you they could not reach you.
You do not find faults like that by waiting for them to show up. You find them because somebody goes looking while nothing suggests anything is wrong.
Faults That Arrive Later#
A website has things with an expiry date. They renew themselves, and usually that works.
When one of those automatic renewals falls out of step, it tells nobody. It keeps trying, day after day, failing, and everything looks normal. Until the day the deadline actually lands. Then your site no longer opens for visitors. Not slowly, not with a warning. It does not open.
Months can pass between the moment the fault appears and the moment you notice. Checking on launch day means checking the one day when everything still works.
And there is a trap of reasoning in here that rarely occurs to anyone. That something is valid today does not prove the renewal works. It only proves it worked last time. Check the state instead of the process and you are checking the wrong thing.
Assumptions That Look Like Facts#
Building involves constant assumption. That is normal and mostly harmless. These photos are all the same size. This page is called what its heading says. It probably runs for other people roughly like it runs for me.
The problem is that a finished assumption looks exactly like a verified fact. You cannot tell from the result whether somebody measured or whether somebody guessed.
We once measured a whole set of photos that had a single size recorded for them, because that is the kind of thing one assumes. Some had different dimensions, several were portrait rather than landscape. None of it was visible anywhere. It only came out when somebody opened the files instead of believing the label beside them.
The unpleasant variant: a check can enshrine an assumption too. We had a test that dutifully verified a particular value, comparing it against a number somebody had guessed beforehand. It passed, every time. It did not find the fault, it confirmed it.
Questions Nobody Asks#
The last point is the quietest and probably the most expensive.
An AI can answer questions and ask follow-up questions too. That does not amount to a complete review. If you do not know about a particular registration, mandatory disclosure or relevant directory, you may leave it out of the task. Without suitable review criteria and access to current sources, the AI can miss it too. An important question then remains open even though the draft looks finished.
That is why a self built site often fails on an absence rather than a fault. Nothing is broken. Something is missing that nobody knew was missing.
A tool can help with that review. Someone still has to define its scope, check the evidence and decide which unresolved points must be settled before launch.
What Follows From This#
None of this argues against AI, we work with it all day ourselves. It handled the part that can be handed to it, and it handled it well.
But all four patterns have one thing in common. The fault cannot be seen in the finished result and it does not report itself. It becomes visible when somebody looks who knows where to look and when. That is not a question of care while building. It is a different activity, at a different time, from a different direction.
Which is why "AI can build a website" is true and misleading at the same time. It can build the website. Whether it works is another question, and no claim answers that. Only a check does.
The next part is about where this costs the most money. The site runs, it looks good, it was done properly, and still nobody finds it.
