Why every AI-generated website looks the same

It is tempting to describe this as a taste problem — the model has seen a billion landing pages and regresses to their mean. That is part of it, but it is not the mechanism, and it leads people to the wrong fix (better prompts). The mechanism is more boring and much more actionable: a coding agent asked to build UI reaches for a component library, and that library ships defaults. Those defaults were designed to be inoffensive, which is exactly what makes them universal.

Ask five different tools — Claude Code, Cursor, v0, Lovable, Bolt — for a dashboard, with no other direction. You will get five layouts and one visual identity. They are not copying each other. They are each independently doing the sensible thing with the same starting material.

shadcn/ui's default theme is built on Tailwind's zinc scale, with slate and gray as the common swaps. It is a cool, slightly blue-tinted grey. Once you have noticed it you cannot stop noticing it: it is the background of most AI-built products shipped in the last two years. It is a genuinely good neutral, which is the problem — it is good enough that nobody replaces it.

The giveaway is not the grey itself but its *uniformity*. A designed interface usually biases its neutrals slightly toward the brand hue, so the greys feel related to the accent. A default interface has a pure ramp sitting next to an unrelated accent, and the two never quite belong to each other.

Inter is an excellent UI typeface, which is why it is everywhere and why it now reads as a default rather than a decision. Where it is not literally Inter, it is the font-sans system stack, which lands on a near-identical grotesque on most machines. Either way the page has one family doing every job: headings, body, labels, numbers.

One family is the actual tell, more than which family it is. A designed system almost always pairs — a display face with character for headings against a workhorse for body, or at minimum a deliberate weight and width contrast within one superfamily. Single-family pages read flat even when every other decision is sound.

shadcn's default --radius is 0.5rem, and it propagates to buttons, inputs, cards, dialogs and popovers alike. Uniform radius across every element size is not how designed systems behave: a 44px button and a 400px card carrying the identical corner radius makes the card look soft and the button look blobby, because radius reads relative to the element it is on.

:root {
  /* The line most AI-built projects never touch */
  --radius: 0.5rem;
}

Hero, then three equal cards, then a testimonial band, then a CTA. This one is genuinely training-data gravity rather than a config default — it is the single most common marketing-page structure on the web, so it is the shape a model produces when the brief does not imply another one. It is not wrong. It is simply the layout that carries no information about your product.

The common response is to write a longer prompt: *use a warm palette, pair a serif display with a grotesque body, vary the radius.* That works. It works for one screen, and it usually works for the second and third.

Then it degrades, and it degrades in a specific way. The agent is not ignoring your instruction — it is holding an increasingly long conversation in which your styling paragraph is competing with everything since. By screen five it remembers *warm* and has lost the hex values, so it approximates. The approximation is a slightly different warm. Screen six approximates the approximation.

People in the middle of this call it *design drift*, and it is the reason a vibe-coded app that looked sharp on Friday looks assembled by committee on Monday. The instruction lived in the conversation, and conversations decay.

A design system stops drift for the same reason it stops it on human teams: the values live somewhere durable and get consulted, rather than being carried in someone's head. For a coding agent, that somewhere is a DESIGN.md plus a token set in the repo. Both are re-read at the start of every session, so screen twenty resolves the same --primary as screen one — not a similar one.

This is the whole argument for giving an agent a real system rather than a better description of one. The semantic token layer matters more than the palette: --primary and --muted-foreground survive a redesign, whereas blue-600 scattered through markup does not.

Mostly not, and that is the more uncomfortable finding. We sampled 299 public DESIGN.md files from GitHub — the file whose entire job is to stop this — and measured what was in them.

Of the 72 that describe a visual design system, 21% specify a radius scale and 26% specify elevation. Those are two of the four tells above, left at their defaults by three quarters of the people who sat down specifically to write a design contract. 69% omit dark mode entirely, so the whole second theme is left to the agent to invent. And 86% name colours by hue or list bare hexes rather than using semantic roles, which means the file cannot be applied to any component it did not explicitly describe.

There is a second-order version of the same problem: 54% of files contain at least one vague adjective standing in for a value, led by *clean* (39%) and *modern* (36%). An adjective does not replace a default. It asks the agent to pick one.

1. 1

   Sample the background

   Pull the hex of your page background. If it is #09090b, #18181b, #0f172a or #020617, you are on an untouched zinc or slate ramp.

2. 2

   Count type families

   Open the computed styles on an H1 and on a body paragraph. Same font-family on both means one family is doing every job.

3. 3

   Compare two radii

   Measure the corner radius on a button and on your largest card. Identical values mean nobody scaled radius to element size.

4. 4

   Diff your first and last screen

   Put the earliest screen you built next to the most recent. List every value that disagrees — that list is your drift.

5. 5

   Decide where the answer lives

   For each disagreement, ask whether the correct value exists anywhere durable. If the honest answer is 'in the prompt I wrote in March', that is the thing to fix.