Machine-made websites converge on each other
I gave five AI tools an identical brief, then measured 945 award-winning websites on fourteen dimensions to see where the machine output landed. It landed in the middle — and, more tellingly, it landed together.
in normalised
feature space
Three different companies. Three different models. Three different technology stacks. Given the same brief, they produced pages that sit closer to one another than two human-designed award-winners typically sit.
That is what convergence looks like when you measure it rather than assert it. The tools are not copying each other. They are each independently walking to the middle of the same room.
sites, projected
to two axes
Every dot is a real website pulled from Awwwards, Httpster, One Page Love, Land-book, recent.design and SaaS Landing Page, loaded in a real browser and measured after its fonts finished loading. The three marked points are AI output.
the centre of
human practice
If machine output were obviously bad, it would sit far from the centre. It doesn't. Two of the three are more typical of award-winning web design than the median award-winner.
| Subject | z-distance |
|---|---|
| Claude Design | 2.16 |
| Replit | 2.34 |
| Median human site | 3.01 |
| Codex | 3.30 |
| 90th percentile human site | 4.31 |
| Most unusual human site | 24.65 |
The machine output is not bad. It is central. Which is exactly what a process that predicts the most likely next thing should produce.
no style words,
first output only
Every tool received this and nothing else. No "modern", no "clean", no "premium" — those words steer every model to the same place and would have manufactured the result.
The page needs to cover: what the pencil is, how it's made, why it costs what it costs, the six-week wait, and how to order.
There's a 15-year-old daughter who does the packaging illustrations. Mention that somewhere.
Real copy, not lorem ipsum. Desktop layout.
The brief was chosen so that nothing in it could be solved from a template. The six-week wait and the daughter are the parts a system has to make a judgement about.
three of the
five tools
Nobody used Inter. Nobody made a purple gradient. Nobody shipped the rounded-card grid. The 2026 checklist for spotting AI design is already out of date.
Instead, four of five independently produced a warm brass accent on a warm ground with a serif display face and uppercase letterspaced labels — two of them landing on the same typeface, Fraunces, with no coordination.
output actually
differs
Two measurements separated the AI output from human practice, and neither is about taste.
| Measure | Human median | AI output |
|---|---|---|
| Largest corner radius | 50px | 0px · 0th pct |
| Grid adherence | 0.62 | 0.78–0.92 · 70–82nd pct |
| Distinct type sizes | 8 | 10–13 |
| Gradient elements | 0 | 0–3 |
Perfectly sharp corners, and a grid it refuses to break. The most expressive cluster of human sites has the lowest grid adherence in the whole dataset — those designers deliberately break their own columns. The machine never does.
Everything else — type scale, palette size, spacing rhythm, density — is indistinguishable.
measured and
how
- 945 sites harvested from six public galleries by following their outbound links; no screenshots taken, no paywalled sources used.
- Each loaded in headless Chromium at 1440×900, measured only after
document.fonts.ready— measuring earlier returns fallback-font metrics and quietly corrupts every type number. - Fourteen features per site: type-size distribution weighted by how much text uses each size, colour weighted by painted area, section-padding discipline, radius, grid adherence, above-fold density, media counts.
- Features winsorised at the 1st and 99th percentile and log-compressed where skewed. Before
this correction, twelve of fourteen features looked bimodal; afterwards only five did. The
difference was a handful of sites using
border-radius:9999999px. - Control: clusters were checked against the source gallery. Purity ranged 0.29–0.35, so the clusters are not simply reproducing which gallery a site came from. Had they, this would measure curation rather than design.
Limitations. Three of five AI outputs could be measured — one produced only screenshots, one only pasted source. One brief, one industry, one viewport. The convergence result rests on three pairwise distances, not thirty.
trying to build
the outputs
Replit's export does not build outside Replit. It ships catalog: and
workspace:* version specifiers that only resolve inside their monorepo, imports a
private package that isn't on npm and isn't used by any source file, extends a
tsconfig two directories above the zip, and requires a Replit-only Vite plugin.
It also imports two hero images it generated — hero-no4-pencil.jpg and
lathe-workshop.jpg — and ships neither.
It also shipped 55 UI components for a static one-page site, including a carousel, a calendar and a one-time-password input. None are used.