Why hire a web designer when AI can build your website?
Building a website with AI takes an hour and costs next to nothing. That is real, and it is impressive. But a website that looks good is not yet a good website. This is what sits in between, and when you can happily do it yourself.

A fair question
You type a few sentences into an AI tool and an hour later there is a website. A big headline, three tidy blocks, a contact form and copy that sounds surprisingly reasonable. So the question is a fair one: why would you still pay thousands to have one made?
We ask ourselves the same thing, because we build with AI every day. It can do a lot by now: produce a first design, write working code, make layouts fit every screen, suggest copy, build components, set up integrations and write tests. For a personal page, a prototype or an idea you want to try out, that is often simply enough.
The difference lies elsewhere. A business website is not a picture that looks right. It is a system that has to work: for visitors, for Google, for your inbox and on the day something goes wrong. That is what this article is about.
AI makes the outside. Who sets the direction?
An AI tool answers the question you ask. "Make a website for a physiotherapy practice" gives you a website for a physiotherapy practice. Not for yours.
The questions that make the difference come before that. Who exactly is this for? What sets you apart from the three others a visitor looks at in the same fifteen minutes? What doubt does someone have just before they call? What should a visitor do on this page, and what is that worth to your business: an appointment, a quote request, an order?
It sounds like groundwork, and it is. But it decides everything that follows: which pages exist, what goes at the top, which proof you show and which button earns its place. An hour of good questions is often worth more than a day of building. That is why our projects start with digital strategy and not with a design.
AI can help with that thinking, as a sparring partner. But it does not know your customers, has not sat in your sales conversations and has no idea which jobs you would rather stop taking.
The problem with AI slop
Spend an afternoon looking at AI-generated websites and you start to see patterns. The same opening with a big headline and two buttons. A purple-to-blue gradient in the background. Rounded cards in a bento grid. A floating 3D object that is about nothing. A serif typeface with one word in italics, for elegance. And a headline like "Unlock your potential", which suits a dentist as well as a software company.
None of those choices is wrong in itself. A bento grid can work very well, and so can a serif. It goes wrong when nobody made the choice. A language model gives you the most likely answer, and the most likely design is the one that already exists most often. That is where the term AI slop comes from: work that is technically fine and belongs to no one in particular.
For a brand, that is a business problem, not a matter of taste. If your site looks like the default output of a tool, nobody remembers it and any competitor can have the same thing tomorrow. An identity of your own works the other way round. It comes from choices that belong to your business: your own photography, a tone that sounds like you talk, a colour with a reason, motion that explains something rather than decorating it. A template can be copied in an afternoon. A brand that stands for something cannot. You can see how different that looks per business in our work.
A quick test for your own site: take the logo away. Could the page still belong to someone else?
Security: the foundation you cannot see
This is where the gap between generating and building is widest, precisely because you cannot see it from the front. A polished interface tells you nothing about what happens behind it.
First, the nuance. Code written by AI is not unsafe by definition, and code written by people is not safe by definition. Sloppy programming existed long before language models did. The risk appears when building goes faster than checking. And building now goes very fast.
Who is allowed to see what?
A made-up example. A business owner has an AI tool build a booking platform. Customers log in, see their appointments and can move them. Everything works and it looks professional. Except the server never checks whether the booking you request is actually yours. Anyone who raises the number in the address bar by one reads someone else's appointment, name and phone number included. Nothing in the interface gives that away.
This is not an exotic scenario. In the OWASP Top 10:2025, the standard list of the most serious risks for web applications, Broken Access Control is number one. According to OWASP, every application in the dataset it examined had some form of it. The organisation is also clear about where the check belongs: in code on the server, where a visitor cannot reach it.
Where are the keys?
A website that sends email, takes payments or calls an AI model needs secret keys to do so. Those belong on the server. If they end up in the code that goes to the browser, or in a public repository, anyone can use them, on your bill.
GitGuardian scans public code on GitHub and counted 28.65 million new hardcoded secrets in public commits in 2025, 34 percent more than the year before. Leaked keys for AI services rose by 81 percent.
What do those figures measure? The number of times a password, key or token was found readable in public code. Not the number of break-ins, and not which share of it is caused by AI: far more people started publishing code in that same year. The report does say that commits co-authored by the AI assistant Claude Code contained a secret roughly twice as often as average, and adds that the human factor remains critical. We work with that tool. So we read that sentence as a working instruction: the leak is rarely in the model and almost always in what nobody checked afterwards.
Does it work, or is it also safe?
A third source points the same way. Veracode had more than a hundred language models carry out coding tasks. In 45 percent of cases the code contained a known vulnerability, and newer or larger models did no better on that point. It is a test with well-defined tasks, not a measurement of real websites. But it shows that "it works" and "it is safe" are two different checks.
What comes with it
For a site that does more than show text, this is the baseline:
- Secret keys live on the server only and never in the repository. OWASP has a practical guide for this.
- On every request the server checks who you are and whether you may see or change this.
- Database permissions are set per user, not to "everyone may do everything".
- Forms validate their input on the server, not only in the browser.
- There is a limit on the number of requests, certainly for a public AI feature. Without one, someone else can have questions answered at your expense.
- Third-party software is kept up to date and security updates are installed.
- There are backups, and someone has tested once that restoring them works.
- There is monitoring, and it is clear who acts when something goes wrong.
None of this is out of reach with AI. You can ask a model to go through exactly this list, and we do. But someone has to know the question needs asking, and be able to judge the answer.
Being found takes more than a meta title
Many AI tools deliver a page with a neat title, a description and a Lighthouse SEO score of 100. That number is a technical check. It says a few basics are in order. It says nothing about rankings, visitors or customers.
Being found starts with search intent: what does someone type, and what do they want to know or do? That tells you which pages you need and how they link to each other. Then comes the technical side. Can a search engine read the content without running JavaScript first? Does every page have one address, with a correct canonical? Is there no stray noindex anywhere? Is the sitemap right? Does it work on a phone?
If you are replacing an existing site, add the question of whether every old address redirects to the right new one. A forgotten redirect costs you the rankings you had. For Harvest Deli we therefore moved the product pages to a clean structure with permanent redirects from the old addresses.
Then the real work starts: content that answers real questions. Google is matter-of-fact about this. It says it looks at the quality of content rather than how it was produced, and states that appropriate use of AI is not against its guidelines. What is not allowed is generating pages with the primary purpose of manipulating search results. So AI is not the problem. Copy that tells nobody anything is.
And then it is not finished. Google Search Console shows which searches you appear for and where you sit just off the first page. If nobody looks at that after launch, nobody knows whether it works.
Fast and usable, on real devices
A site is usually built on a fast laptop with a large screen and good wifi. Your visitors are on a train, with a four-year-old phone.
Google measures the experience of real users with the Core Web Vitals: is the largest element on screen within 2.5 seconds, does the page respond to a tap within 200 milliseconds, and does the layout stay put while loading? Generated sites often trip over the same things here: heavy images, a lot of JavaScript for animations and fonts that load late.
Accessibility is the other half. Can you operate the menu and the form with a keyboard alone? Is the text readable on a phone in sunlight? Does a screen reader understand what a button does? That starts with semantic HTML: a button is a button, a heading is a heading. It is work you cannot read off a screenshot, so it is the first to be skipped. You test it by doing it, in several browsers and on real devices.
Who actually owns it?
Ask the following about any website, whoever or whatever built it.
- In whose name is the domain registered?
- Where is the code, and do you have access to it?
- Where is your customers' data, and who can reach it?
- Which external services run alongside: email, payments, analytics, AI?
- Which cookies are set, and only after consent?
- What does it cost per month in hosting, subscriptions and maintenance?
- Can you leave? What happens if the builder stops or the platform gets more expensive?
With an AI builder that keeps everything in one environment, that last question is the one to ask up front. Some let you take the code with you, others do not. Ask an agency exactly the same thing. Anyone who cannot explain how you would leave them is not a partner but a dependency.
Launch is the beginning
The day a site goes live feels like the end of the project. In practice, it is the start of the part that decides whether it pays off. Does an enquiry through the form really arrive, three months from now too? Is the confirmation email delivered, or does it land in spam? Which pages lead to enquiries and which do not? Who installs the security update that comes out in February? Who notices when the site is down on a Saturday evening?
A form that breaks silently is the classic example. The page looks fine, only nothing comes in any more. You find out when someone calls to ask why you never replied.
Responsibility after launch is less visible than a design and worth at least as much. We call it ongoing growth: measuring, maintaining and improving based on what visitors actually do.
What sits behind our own site
In early November 2026 we replaced our own site. AI was our most important tool in doing so, which is exactly why it is a useful example of the work you do not see.
- More than four hundred redirects take old addresses to their new place, so existing links and search results keep working.
- Before the switch, all 114 pages were checked for status, title, a single H1, canonical and indexability, and every internal link for broken targets.
- The contact form checks every enquiry on the server and stops automated submissions.
- The AI assistant on the site talks to the model through our own server. The key never reaches the browser.
- Analytics only load after a visitor has given consent.
- Every release comes from reviewed code, is tested on a separate address first and only then promoted. The previous version stays ready to roll back to.
- Hundreds of automated tests run in several browsers before anything goes live.
Not everything is finished. Google is still processing the new addresses, and we follow that in Search Console. The monitoring that checks the site and the form independently of our own computers is being set up right now. That is part of it: a website is a process, not a delivery.
This is how we work for clients too: strategy, design, development, security, testing, launch, measuring, improving. AI is in almost every step. It replaces none of them.
When building it yourself with AI does make sense
Not every business needs custom work, and not every website needs an agency.
Building it yourself with AI is a fine choice when:
- it is a personal project or a small informational site;
- you want to test an idea before investing in it;
- there are no payments, accounts or sensitive data involved;
- you have enough technical knowledge to check the result, or know someone who will.
Professional help is worth considering when:
- the site is an important source of new customers;
- it involves payments, personal data or customer accounts, as with an online shop or a customer portal;
- you connect several systems, such as a calendar, stock or accounting;
- you are replacing an existing site that already ranks in Google;
- a business process depends on it;
- your brand needs to stand out visibly.
There is a middle road as well. Build a first version yourself with AI and have it reviewed for security, search and speed. Or start with a compact professional site and expand once it works. For a rough idea of cost, read what a website costs. If you are unsure whether your current site is good enough, look at these five signs your website is costing you customers.
Generating is not the same as building
The question was why you would still have a website made when AI can do it. The honest answer: for the outside, you need to less and less often. For everything that turns a website into a working part of your business, little changes. Someone has to choose the direction, give the design a face of its own, lock down the back end, check that Google and visitors can work with it, and still be there when something goes wrong.
That does not have to be an agency. It does have to be someone.
AI has made making a website easier. Making a good website is still a craft. If you want to know what that means for your situation, see how we approach web design and how we use AI, or book an introduction.
