The sites clients show me
More and more clients have come to me saying that they can use AI to make their website. They're right that they can, and the tools are good enough now that the result usually works: it loads, it's responsive, and the contact form sends. It ticks all the boxes from the perspective of someone who just wants a website online. What strikes me when they show me what they've made, though, is that the sites all look the same.
That isn't a complaint about AI, which I use for certain parts of my own work. It's an observation about what happens when nobody makes the design decisions, and about what that costs a brand whose whole job is to be recognisable.
What "the same" looks like
The resemblance is specific enough to list. There are eyebrows everywhere, which are the small uppercase labels that sit above each heading and announce what the section is about before the heading says it again. There's monospace text, the kind of type designed for code, used on labels and dates to make a site look technical whether or not it has anything to do with software. The body copy is set in Inter, a well-made typeface that has become so common it now reads as the absence of a choice. I do personally still like Inter, and unfortunately due to seeing it on every AI generated design I have to be more careful in my usage.
The colours come in two flavours. One is the purple-to-blue gradient that has sat behind a great many startup headlines over the last few years. The other is its more recent replacement, a warm beige with a terracotta or muted orange accent, which tends to appear when someone has asked the tool for something that looks less like AI. Underneath either palette you'll usually find the same structure: a centred headline, a single button, and a row of three rounded cards, each with an icon and a sentence.
None of these things is wrong on its own, and plenty of good sites use one or two of them deliberately. The problem is that they arrive together, on sites for an accountant, a candle shop and a logistics firm, with nothing to tell the three apart. Each part of a design should be doing something for the user, either helping them move through the website or telling them something. Nondescript icons on every card don’t add anything, and even from the design side they’re a problem, because they look as though they’re trying to say something when they aren’t.
Why it happens
An AI model doesn't design a website so much as predict one. Given a prompt like "a modern landing page for my business", it produces the most likely answer, and the most likely answer is an average of the enormous amount of front-end code it learned from. A large share of that code was built with the same few frameworks and starter templates, which means it shares their default fonts, colours and component shapes. The model isn't choosing Inter or the purple gradient because they suit your business. It's choosing them because they're what a website most often looked like in the material it learned from.
This is also why asking for something "less AI-looking" doesn't solve much. The model hasn't learned what your brand is, only that the purple version is now considered a cliché, so it reaches for the next most likely alternative. That's where the beige and terracotta comes from, and designers who track these patterns describe exactly this: ask for a less AI-looking page and the cream and terracotta arrive in its place. A tool working from an average can only ever move you to a different average.
Much of the sameness traces back to defaults in popular tooling. Tailwind CSS's indigo and violet shades, Inter as the font in most starter kits, and component libraries such as shadcn/ui with their particular corner radius and card styles all appear constantly in the code these models were trained on. A common explanation among developers is that the defaults were everywhere because most people never changed them, so the model learned them as the norm. Specifying a design system up front helps, but somebody still has to design that system. Once a design system is in place along with a nice library of components designed bespoke for your brand, using AI to brainstorm new component ideas massively improves.
Familiar isn't the problem
It would be convenient to argue that people dislike generic design, but the research points somewhere more interesting. In a 2012 study by researchers at the University of Basel and Google, participants rated screenshots of real websites after seeing them for as little as 17 milliseconds. The sites they found most appealing were those that looked simple and looked typical of their category, and as the researchers summarised it, an unfamiliar design was judged less attractive even when it was simple.
That goes a long way towards explaining why AI-built sites seem fine at first. They're uncluttered, and they follow the conventions people expect, so they pass the first glance. A good designer works within those conventions too, because a navigation bar people can't find helps nobody.
The difference is in everything the conventions leave open. Layout conventions tell you where the menu goes, but they say nothing about the typeface, the colours, the photography, the way a section opens or what the site sounds like, and those are the decisions that make a business recognisable. On an AI-built site they've all been filled in with the default, which is why the site can look perfectly reasonable and still look like nobody in particular.
Why customers notice
Research led by Gitte Lindgaard at Carleton University found that visitors form a judgement of a site's visual appeal within 50 milliseconds, and later work has found that this first impression goes on to colour how people rate the site's usability and trustworthiness. Whatever a visitor concludes in that first moment, the rest of the site has to argue with.
There's also some evidence that people are growing warier of anything that feels machine-made. In a survey published by Kinsta in September 2026, 58% of consumers said they trust content less when they believe it was AI-generated, and 61.1% said they associate the quality of a company's website with the quality of its products or services. It's a US survey commissioned by a hosting company, so I'd treat it as a sign of the mood rather than proof, but it sits comfortably alongside the first-impression research.
I don't think most customers could tell you that a site uses Inter or that its palette came from a framework's defaults, and they don't need to. They register that they've seen this site before, on a dozen other businesses, and they draw the reasonable conclusion that nobody here spent much time deciding how the company should come across. For a brand, that's an expensive impression to give, because your website is the one place where you decide entirely what people see, and a default look gives that decision away.
Where I use AI, and where I don't
I should be clear about my own position, because this isn't an argument against the technology. I use AI when developing a site, as it's very useful for boilerplate code and functionality, and it saves my clients time and money on work that was never where the value lay. For design, however, I don't use AI at all, and that's a promise.
Instead I spend a long time researching competitors' sites, trying several different designs, and thinking about the customer journey and how each of these decisions affects SEO and customer retention. Most of that work never shows on the finished site, which is rather the point, because what shows is the result of having made the choices deliberately.
Two recent projects show what that looks like in practice. For Motion Assembly, an architecture firm whose value lies in 3D scanning and modelling, we didn't list the process in text and images but built a section where scrolling moves a video through each stage of a project. A video that scrubbed continuously with the scroll would have been too heavy for older laptops and weaker connections, so each scroll moves to the next stage and pauses, and visitors can switch to a slideshow if they need to. The blog is laid out as a book that readers page through, whilst the underlying HTML stays plain enough to read well for search engines.
For Wabital, a social media support company working with clients in and outside Japan, the research began with Japanese and western sites in the same space. Most of the Japanese competitors leaned towards a corporate, tech-forward look that didn't suit Wabital, so we kept those influences for the functional parts of the site and built its own identity around the brand-led sections, with a bento-box layout for services and vertical Japanese text alongside each product card. No prompt would have produced either of those sites, because each came from a question about one particular business and the people it needed to reach.
When an AI-built site is enough
There are times when a generated site is a perfectly sensible choice. If you're testing an idea before committing to it, or you need a single page online by Friday so that people can find you, a quick site that looks like many others does the job, and it's better than nothing. Even Motion Assembly started with a simple landing page, on the understanding that we'd build the full site once the business knew what it wanted to say.
The cost arrives later, once the site is meant to carry the brand: when it's the thing a prospective client looks at before getting in touch, or the place a customer goes to check you're the real company they heard about. At that point a default look is no longer saving you money. It's quietly telling people that you're interchangeable with everyone else who used the same tool.
If your site is already at that stage, it doesn't usually need to be thrown away. Much of the content and structure can stay, and a redesign starts from the decisions that were skipped the first time: who the site is for, what they should feel when it loads, and what would make it unmistakably yours.
