Start a Project
Menu
Go back

Every AI website looks the same. It's costing you customers

Bhaskar Bhandari
Bhaskar Bhandari, Founder, Nord Studio

8 min read

A classical marble bust in gold sunglasses sipping an iced coffee, ringed by a carousel of near-identical yellow AI landing pages.

Your site looks good. So does everyone else’s, and that’s the problem.

If you built it this year with an AI tool, I can probably describe it without opening it. Big centred headline. Inter, or something a shade off Inter. There’s a soft gradient behind the hero, three cards with rounded corners and a small icon in each, and a pricing table with the middle tier highlighted. It’s clean, it loads fast, and the person looking at it has seen four of them already this week.

People have a name for it now. AI slop.

This post is for the people shipping that site. Founders, solo builders, small teams without a designer, who reasonably assumed sometime last year that design had been handled.

We use AI here every day, on this site included, so take the rest as a working note rather than a complaint about robots.

Easy stops being an advantage the moment it’s easy for everybody

The distance between having an idea and having a working product shrinks every year, and 2027 will shrink it again. That’s good. Fewer people get stuck because they can’t wire up a contact form.

But a capability everyone has isn’t a capability. It’s the floor. Nobody advertises that their shop has electricity.

So when a founder tells me their generated landing page looks premium, what I’m looking at is the current default of a tool that ten thousand other people opened the same week. Premium is a comparison. It needs something worse to sit next to. Take that away and the word does nothing.

The tools converge because averaging is what they do

Doshi and Hauser gave writers story ideas from a language model and measured what came out. The individual results improved. Stories written with AI help scored higher on creativity, on quality of writing, on how much people enjoyed them, and the weakest writers gained the most. That’s the democratisation everyone promises, and it showed up.

Then they measured the stories against each other. The AI-assisted ones were more similar to one another than the human-only ones were. Better on their own, more alike as a set. The authors’ worry was that an industry all drawing from the same well ends up with a narrower output than it started with.

Now put landing pages where the stories were.

Every builder is trained on roughly the same library of good design, so every builder lands on roughly the same average of good design. No single product did anything wrong. That’s just what averaging produces.

And these tools are good now, which makes it worse rather than better. A benchmark called UI-Bench ranked ten AI text-to-app tools using 30 prompts, 300 generated sites and more than 4,000 expert pairwise comparisons. Who won is beside the point. The field needed a leaderboard because the outputs had gotten close enough that you couldn’t separate them by eye. When everyone clears the bar, the bar stops being where you compete.

Nobody sees your backend. Everybody sees your interface.

Most arguments about AI and craft get run as a single question. There are two, and they come out opposite.

Your backend is private. If a model wrote it, if you lifted it from a gist, if it’s held together with something you’d rather not explain in an interview, no customer will ever find out. Your API can fall over on a Tuesday and people will be annoyed about the outage, not about who authored the code that caused it. I’ve never heard anyone look at a slow endpoint and say that looks AI-generated.

The interface gets no such privacy. For anyone who hasn’t signed up yet, it is the product. Your interface is what gets screenshotted and sent to a friend, or opened in the tab beside a competitor. Sameness is visible there and nowhere else, and people use what they can see to guess how much care went into what they can’t.

That’s the asymmetry. Generated code is invisible, so it costs you nothing. Generated design is the storefront.

What the eye catches

People can spot an AI-built site, and they got quick about it. What they’re reacting to is narrower than the gradient, though.

Look at a page that somebody designed and you can reverse-engineer the arguments they had. This heading is enormous because the business lives or dies on one action. That section is small and grey because it matters to four customers a year and cutting it entirely would have been worse. Somebody put the testimonial under the price, having watched people hesitate at exactly that spot. Every one of those is a decision, and decisions leave marks.

Generated pages come out flat. They’re competently made, but flat in priority. Every section gets the same padding and the same confidence, because nothing was ever ranked against anything else. Uniform polish reads as an absence, and the absence is judgment. A page where everything is equally important is a page making no argument at all.

Slop is a fair word for what comes out, even if nothing in it was made carelessly. Deciding which parts deserve more attention is the step that got skipped.

The font was never what gave it away.

And it shows up in the trust numbers. A CHI 2026 study put 43 people in front of six websites, some using real photography and some using AI-generated images, and had them judge twice, first blind and then knowing which was which. Blind, they leaned toward the AI imagery. It looked better. Once they were told, trust swung back to the real photographs, hardest in health and government contexts. Participants described the tension themselves, appeal on one side and authenticity on the other.

The AI version won on looks and lost on belief. In 2026 you don’t need a disclosure label for that to happen, because people are doing the disclosing themselves, in about a second, from the layout.

What half your traffic decides in 50 milliseconds

None of this kills a good product. If yours solves someone’s problem, a generic site won’t sink it. Nobody walks away from a tool that works because the corner radius is unfashionable, and I’d take a useful ugly product over a beautiful useless one every time.

Usefulness only counts once somebody stays long enough to find it.

Lindgaard and her colleagues found that people form a first impression of a web page in roughly 50 milliseconds. You can’t read anything in 50 milliseconds. They’re reacting to a shape. And the Stanford Web Credibility Project found that 46.1% of people gave visual design, the layout and typography and colour, as a reason they did or didn’t believe a site, climbing past half for finance and travel.

Close to half your traffic is deciding whether to trust you off the shape of the page. If that shape is identical to fifteen others they’ve seen, you’ve handed the decision to a coin. That’s what it costs when your site looks the same as everybody else’s, and the bill gets paid by the people who leave before they read a sentence.

Direction is the input, and it’s the part you have to bring

Almost everyone complaining that AI design looks generic is prompting generically. “Make me a modern SaaS landing page” is a request for the average. The average is what comes back. The model did its job.

What changes the output is bringing something to it. A reference that isn’t already in every training set. A palette that isn’t one of the six safe ones. A constraint from the business itself, like the fact that your buyers are procurement managers who need a PDF, not developers who want a code sample. An opinion about what should be loud and what should shut up.

None of that comes out of the tool. Somebody has to hold it and keep reapplying it, because the model drifts back to the median every time you look away. Give it direction and it’s the fastest execution partner anyone has ever had. Give it a sentence and it’s an efficient way to produce something forgettable.

Most of what we get hired for at Nord starts before anything gets drawn. Someone has to decide what a page is for, what the single action is, and what gets sacrificed so that action is impossible to miss. The drawing part is quick now. The deciding part is where the work went.

2026 lets it slide. 2027 won’t.

You can still get away with a lot. In plenty of markets a decent generated page clears the bar because the competition is also a decent generated page.

That window is closing, and closing for a boring reason. More models next year, better at UI and at images, and every one of them lifts the floor. When the floor comes up, everything standing on it flattens into one silhouette, and the only thing with an outline is whatever somebody deliberately pushed higher.

There’s a practical wrinkle too, and it’s why the gap won’t close on its own. Doing this properly is slow. Collecting references and arguing about direction until the thing has a personality takes hours, and most people want the page live by Thursday. People ship generic because they’re in a hurry, not because they couldn’t do better.

Which is the good news if you design things. Machines produce a nice screen for a few cents now, and they’ll produce a nicer one next year. The remaining skill is knowing which nice screen is the right one for this particular business, and being willing to spend the afternoon that the shortcut skips.

Sources

If your product works and the site in front of it looks like everyone else’s, that’s a direction problem rather than a tooling one. We design and build web, mobile and AI products at Nord, and most of that job now is making the calls a prompt can’t make for you.

Book a call about your site. We’ll tell you which parts read as generic, and we’ll say so if none of them do.

Have something you need built?

Tell us what you're making. You'll get a straight answer on scope and timeline.

Start a Project

Related posts