Should You Just Build Your Website With AI?

Ask an AI for a landing page and you will have something on screen before your coffee goes cold. It will look fine. It will have a hero, three benefit blocks, a testimonial and a form. Then someone quotes you thousands of dollars for what appears to be the same thing, and the quote looks like a joke.

That reaction is reasonable, and we are not going to tell you the tools are bad. We use them every day. They made us faster. If they had not, we would be the ones being replaced.

But we should be straight about what got faster, because it is narrower than it looks. And the honest answer to the question in the title is sometimes, and it depends on what you need it to do.

The saving is smaller than it appears

A genuinely good page still takes us about three days.

That surprises people, so it is worth explaining. What the tools compressed is the typing: turning a decision into markup, drafting copy that is roughly right, producing a layout, filling in the boilerplate every site needs and nobody enjoys. That part used to take days and now takes hours, and the saving there is real.

What they also did is add work that did not exist before.

Prompt a site screen by screen and the output arrives without a system underneath it. Random hex values rather than a palette. Typefaces that do not quite agree with each other. Spacing chosen fresh each time, and each screen locally plausible without the set of them holding together. Pulling all of it onto one consistent system takes longer than generating the screens did.

So the three days did not go away. They moved.

Generic is the default, not the ceiling

That consistency problem is the one people write about, and it is the one currently being solved. Builders tuned for websites already handle much of what raw prompting gets wrong, and the next generation will handle more. If your objection to these tools is that the second page does not match the first, expect to be out of date shortly.

The problem underneath it is not going anywhere.

These tools are converging on competence, and competence is the wrong target if the point is to not look like everyone else using the same tools. Whatever you build with, general model or purpose-built builder, it is optimising toward a well-executed version of the most probable website for your category. The colours will vary. The typography will vary. There may even be a genuinely interesting element or two. It will still read as AI generated to anyone who has seen four other sites in your industry.

The clearest evidence is public. Somebody posts the site they just built, somebody replies that it looks AI, and they are usually right. Not because anything is broken. Because the spacing conventions, the layout decisions, the motion and the hierarchy all came out of the same distribution. Individually defensible. Collectively anonymous.

This has happened before. It is what a WordPress template does to a business, and the feeling is identical: you land on a site, you recognise the theme, and something quietly deflates. The company might be excellent. The site has just told you it is not trying very hard. AI generated is becoming the new recognisable theme, and recognisable within a category is the opposite of differentiated.

That costs more here than it would anywhere else, because your site is the one channel where you are not constrained. Every platform hands you the same grid, the same player, the same aspect ratio as your competitor. The site is the only place you can look like yourself. Settling for the first result spends that advantage for nothing.

So the distinctive version is reachable, and it is not one prompt away.

Say you want a hero that does not look like the four other sites in your industry. A layout convention that carries a particular point, an icon sitting somewhere unexpected, an art element doing actual work rather than decorating. You can get there with these tools. What it takes is design discovery first, so that you know what you are reaching for, and then a run of prompting loops to close the distance. Call it thirty minutes to two hours for one hero that has an edge on it. Motion is the same shape: you can direct a considered flow this way, but not in one shot, because the model has no view on what the movement is meant to communicate.

What you are looking for is a narrow band. Specific enough to feel considered, not so eccentric that it loses people, and positioned to match where the business actually sits in its market. Finding that band takes knowing the vertical well enough to see where its edges are, and knowing the brand well enough to know where it belongs between them.

That is where the remaining expertise lives, and it does not get easier to prompt as the tools improve. The model has seen everything in your category averaged together, and the average is precisely the thing you are trying to step away from. The loop is cheap. The reference point you bring to it is not. Someone without design experience can prompt for six hours, get something polished, and have no way of knowing they stopped short, because nothing in the process tells them when they have arrived.

What you cannot prompt your way through

Not because the tools are weak. Because these decisions depend on knowing things about your business, your customers, and how software is maintained.

Knowing what to leave out. Generated pages are additive. Ask for a landing page and you get every section a landing page has ever had. The hard call is which two your customer actually needs, and what improves when you delete the other five. Subtraction takes a view about the business, and nothing will form that view for you.

Information architecture and user journeys. This is the one most often mistaken for taste. A page for someone comparing five options on price is a different page from one for someone who is frightened and deciding today. The layouts look similar and they perform nothing alike.

Working that out is a process rather than an instinct: establishing who is arriving and where from, what they are trying to accomplish, what the page is structured to make happen, where people currently fall out, and then measuring afterwards to find out whether any of it worked.

AI is genuinely useful inside that process, and we would recommend using it. Ask it to map the common journeys for your vertical and you will get a serviceable starting set in minutes, which beats staring at a blank page. What it cannot do is tell you which of those journeys is yours, because it has no objective. You do, and the structure has to be built backwards from it. An AI will build you something that functions and never mention that nothing is being measured. Not through carelessness. It has nothing to measure against.

Art direction. The overall visual argument: what this should feel like, and why that suits your business rather than the one next door. It is the difference between a page that is competently assembled and one that has a point of view.

Content. Someone has to decide what this business actually says, in its own voice, including the parts specific enough to be worth reading. Drafts come easily. A position does not, and the failure is harder to spot than any of the others on this list, because empty prose reads exactly as well as full prose. We wrote about that separately in what AI writing actually costs you, using this blog as the evidence.

The technical choices. Which framework, which libraries, how the code is organised so that a person can find things in it a year from now. AI will happily pick for you and give no indication whether the choice was sound.

Code quality, which is invisible from the front. A page that renders correctly tells you nothing about what is behind it. Some builders do structure things more coherently under the hood, but none of it is guaranteed, and a polished surface sitting on tangled code looks exactly like a polished surface sitting on good code. The bill arrives later, when somebody needs to change something and finds that the cheapest option is to rebuild the page rather than edit it.

Consistency in the small things. Icons that come from one family, or are drawn for you. Buttons that behave identically everywhere. States that all exist. Individually trivial, collectively the difference between finished and nearly finished.

Quality assurance as a process, not a page. Does the whole path work? Arrive, understand, decide, act, get a confirmation, and land somewhere that makes sense. Deploy it and it renders. Whether it functions as a process is a separate question, and it is the one that decides whether you get customers.

What this looks like in practice

We have a client who has been prototyping with AI, and it has been genuinely good for both of us.

They build the thing they are imagining and hand it over. That is a real gift: instead of a description we have to interpret and guess at, we get something we can open. The back-and-forth where we build our best understanding of an idea and hope it matches what was in their head largely disappears. The feedback loop gets much shorter, and the project moves further, faster. We would encourage it.

Then they ask whether it can be launched, and we look properly.

Emoji used throughout as interface. A hundred and thirty-seven different hex values for green, because there is no palette, just a fresh decision each time something needed to be green. Typography that shifts between screens. Border radius that changes depending on when the component was made. All the ordinary consequences of building without a system, none of which are visible while you are building and all of which are visible afterwards.

And the part that matters more: the experience is uneven. Some paths feel smooth. Others feel convoluted, because a feature was prompted and the model filled in the gap plausibly, without anybody holding the whole flow in mind. Locally sensible, as a journey somewhat strange.

None of that makes the prototype a mistake. It got the project a long way in a short time, and it communicated the intent better than any brief could. But the distance between that and something you can put in front of paying customers is real work of a different kind: consolidating decisions, building the system that should have been underneath it, and walking every path as a user rather than as its author.

The prototype gets you a long way down the field quickly. The last stretch is still the last stretch.

A multiplier, not a replacement

The most accurate thing we can say about these tools is that they multiply whatever expertise is already there. Design, development, copywriting, all of it. Someone with ten years behind them gets faster and often better, because they know what to reject. That is the actual skill now.

Without that, the failure mode is quiet. The result will not look like AI made it. It will look fine. What goes missing is the edge you were reaching for, plus a scattering of small things that an experienced person would have caught and a first-time prompter cannot see yet.

Ask an expert copywriter to look at generated copy and they will find the claim that does not land. Ask a designer to look at a generated poster and they will point at four things in ten seconds. Neither result is bad. Both are ordinary, and ordinary was probably not the goal.

How to tell whether what you built is any good

Most of this is checkable, and none of it needs you to be a developer. The quickest route is to ask the AI that built the thing, because it has the whole codebase in front of it and answers honestly when asked directly, which is a better test than taking our word for it.

The single most useful question is not a technical one. It is what this site is supposed to make happen, and whether every page still points there. Prompted section by section, a site accumulates rather than argues: each part reasonable, the whole no longer saying anything in particular. The model never asks that question because it has no objective to measure against.

We put the rest of it in one place, along with the prompts to run and the standard pre-deploy passes that do not run themselves: the checks worth making before you launch an AI-built site.

So: should you use it?

Genuinely, yes, in a lot of cases. If you need something clear and functional and professional, and "fine" is the right target, the tools will get you there and you will have saved real money. We would tell you so.

But if you need the edge, if you are competing with people who are also competent and the difference is whether your business feels like itself, then you need an expert in the loop. Not to type. To decide, to reject, and to notice the things you cannot see yet.

That is what you are paying for, when you pay for this. Not the page.