Building a website with AI: what works, what does not

The question now comes up in almost every first conversation: do you still need someone to build a website when a model produces a complete page in two minutes? The honest answer is neither “yes, obviously” nor “no, that is a toy”. It depends on which job you hand the tool — and this article tries to draw the line concretely enough that you can check it yourself.

One thing up front, so you can place the rest correctly: we use these tools ourselves. Exactly how is further down. This is not an article talking down a tool because it touches our own work.

What AI genuinely does well today

Start with what works. The biggest benefit is not the finished page but getting away from the blank screen. Reacting to a mediocre draft is far easier than starting from nothing. A text you reject moves you forward more than an empty page you stare at — because rejecting it is the first time anyone says out loud what should have been there instead.

• A first outline: which sections a service page usually has, and in what order they run
• Rough drafts that show you what is missing — you almost never notice that from an outline, only from a concrete sentence
• Five versions of a headline when your own has got stuck
• Alt text for images as a starting point, which you then adjust to the image you are actually using
• Translation drafts, when a second language version is planned and someone checks it afterwards
• Summaries of long notes, for instance from a customer conversation you wrote down

That is not a small benefit. Anyone who has tried to write the copy for their own website on a Friday evening knows the resistance is not in the phrasing, it is in the starting. That resistance is exactly what the tool reliably removes.

What AI cannot do — and why that is the expensive part

A model does not know your prices. It does not know the three objections you take apart on the phone every week. It does not know which of your services actually earns money and which you only still offer out of habit. It does not know your legal obligations, and it does not know which customers you actually want. All of that is a precondition for a page that does anything.

The decisive point sits one level above that: a model does not decide what a page is for. It fills a page. Whether that page should generate enquiries, bring in applications, deter enquiries that do not fit, or simply build trust is a question it cannot answer — and it will not tell you that it failed to answer it.

That is where the real risk comes from. Obviously wrong copy is harmless: it stands out and gets deleted. The expensive kind is copy that reads coherently and is wrong in one detail. Nobody corrects that, because nobody thinks it needs correcting.

An example of how this really happens: a draft writes “We advise you free of charge and without obligation at your premises.” Sounds harmless, sounds like every service business page. But if on-site visits are only free within your immediate area and charged beyond it, your website now carries a promise you have to withdraw in the first phone call. The conversation starts with a disappointment — and you never even noticed where it came from.

Job by job: where the tool holds up

Rather than talking about “AI websites” in general, it helps to look at individual steps. The line does not run between text and technology, it runs between phrasing and committing. The tool can phrase. It cannot commit.

Generic in exactly the places that must be specific

What stands out is where generated copy goes vague. Not on the incidentals — there it writes cleanly and fluently. It goes vague at precisely the four points that make a website useful in the first place:

• Price: “fair and transparent pricing” instead of a figure, or at least the logic that produces one
• Process: “tailored individually to you” instead of what happens in week one and what you have to supply for it
• Commitments: “the highest quality” instead of a statement you can be measured against
• Audience: “for companies of every size” instead of the two customer types you genuinely work best with

That is logical: none of those four exist anywhere online, because they only exist in your head and in your invoices. A model can only produce language out of existing language. The very part it cannot supply is the only part of your website your competitor does not also have.

A one-minute test: take a generated paragraph and delete the company name. If it would sit equally well on three other businesses’ websites, it is not finished. That is not a style problem but an information problem — and it cannot be fixed with better words, only with more knowledge.

Two hard limits: legal texts and images

Legal texts do not belong in the tool

An imprint, a privacy policy and — depending on what you sell — a cancellation notice and terms of business are not content, they are mandatory disclosures. What belongs in them depends on your legal form, whether you are VAT-registered, what professional title you hold, which services on your site process data and whether you sell to consumers. A generated text knows none of that and still looks complete. That is the dangerous combination.

The same applies to the privacy policy in the other direction: it has to describe what your website actually does. If it lists services you do not use, or omits ones that are embedded, the text is not merely useless but wrong. These texts belong with a lawyer or with a maintained, vetted generator — and in both cases someone has to keep them matching what is really embedded.

Generated images do not show your work

With images the issue is less legal and more fundamental. A generated photo of a team is not a photo of your team. A generated piece of work is not one of your pieces of work. Putting that on the site as a reference or as “our team” misrepresents your business, however good the image looks. And visitors now recognise generated imagery far more often than most businesses would like.

On top of that, rights and provenance for generated images are currently unsettled. That is a second reason for caution, though not the main one. We do not use generated images to depict real work, real people or real results. For abstract illustration or backgrounds it is worth discussing; for anything a reader could take as evidence, it is not.

The practical alternative is less convenient and works better: an afternoon with a decent camera in your own premises. Real rooms, real people, real work. It beats any perfect stock photo and any generated motif, because it is the one visual language your competitors cannot also use.

The sequence that actually works

None of this adds up to a ban. It adds up to a sequence. It is unspectacular, but it is the difference between a website that looks like something and one that is true:

1. 01Settle the purpose before anything is written: who lands on this page, with what question, and what should happen next. A person makes that call.
2. 02Have a draft generated — outline and rough text. Explicitly as material, not as a result.
3. 03Review and rebuild the structure: what is missing, what is duplicated, what sits in the wrong place. Usually the most important information moves up.
4. 04Verify every fact: prices, deadlines, services, responsibilities, opening hours, locations, figures, names. Anything you did not supply yourself counts as invented until proven otherwise.
5. 05Rewrite every sentence that makes a claim. Claims are the part you stand behind — they have to come from your business, not from a language model.
6. 06Handle legal texts and images outside this process: legal texts via a lawyer or a vetted generator, images from the real business.
7. 07Name one person who has read everything before it goes live and is responsible for the content. Without that step, nobody decided anything.

Steps four and five are the whole workload — and the point where you find out whether any time saving survives. On standard sections it usually does. On a service page with prices and commitments it is often faster to treat the draft as a memory aid and write the text yourself.

How we work with it ourselves

We use these tools day to day: for first outlines, for rough drafts, for headline variants, for alt-text suggestions and as a second opinion on a paragraph that will not settle. Hiding that would be silly — most readers assume it anyway, and the credibility is worth more than a claim that everything is handwritten.

The rule behind it is simple and has no exceptions: every published statement is a person’s responsibility. Prices, deadlines, commitments and legal texts are not generated. Every figure is checked against its source. And anything we write about a client comes from the conversation with them, not from something the model filled in.

When we rebuilt our own website in August 2026, that was exactly the process. The part that really cost time was not the writing. It was deciding what each page had to achieve, and then checking every single statement on it. The tool made starting easier and did not help with the rest.

So do you still need someone to do this?

Honest answer: not necessarily. If your offering is clear, you enjoy writing, you know your customers well and you are willing to learn a system, a generated draft plus a common site builder or a simple setup will get you further than was possible a few years ago. That is not a sentence against our own offer, it is an observation.

Outside help becomes worthwhile elsewhere: when structure and priorities are unclear, when the site has to sell rather than merely exist, when technology, speed, accessibility and legal texts all have to line up — or simply when your time is worth more somewhere else. The difference is rarely in the writing. It is in the deciding and the checking, and no tool has taken over either of those yet.

Frequently asked questions about AI-built websites