Why every AI-built website looks the same

You know the page before it finishes loading. A purple gradient, a pill that says Powered by AI, a centered title, two buttons, three cards with an emoji on each. It isn't ugly. It's just nobody's.

That isn't a failure of taste, because the model doesn't have any. Ask it for a landing page with no reference and it does the only thing it can: it writes the most likely page. The most likely page is the average of every page it has read.

A screenshot doesn't fix it

The obvious move is to paste a screenshot of a site you like. It helps with layout. But a screenshot is the result, and the model has to guess everything that produced it: the font, the size, the tracking, the exact grey, the radius, what happens on hover.

So it guesses, and you get a page in the neighbourhood. Close enough to see what it was going for, far enough to feel wrong. Most of what makes a design feel considered lives in numbers too small to read off a picture.

What to give it instead

Give it what the browser actually applied. There are two good ways to do that, and they answer different questions.

If you don't know what you want, use a library of shipped designs like Mobbin or Refero. Your agent can search real paywalls and onboarding flows and see what's common.

If you've already seen what you want, measure that page. This is what plagio does: it opens the site in a real browser and returns the markup, the CSS, the design tokens, the hover states and the motion.

Then make it yours

Reading how a good page is built is how people have always learned design. Shipping someone's logo, photos, words or whole page isn't. Those belong to their owners, and a site that's a copy of another site has the same problem as one that's a copy of the average. Take the measurements. Keep your own product.