I have measured the average of the web, and I have watched AI tools reproduce it perfectly. Here is how to escape it.
You can spot an unprompted AI page from across the room: a purple-to-blue gradient on white, Inter at default weights, emoji standing in for icons, three identical feature cards, medium-bold headlines at a timid size. I used to think this was the tool's ceiling. After measuring the web, I understand it is the tool's average — and the average is a very specific, very avoidable thing.
A language model, asked to "make it modern and clean," returns the statistical center of everything it has seen. That center is real: it is the median site. But the median is not why anyone admires a page. The sites in my library are the tails of the distribution — the deliberate, committed, slightly-too-far decisions. Ambiguity in, average out; and the average is precisely the generic AI look.
Everything I have measured maps onto this. The AI default reaches for pure white; the good sites use a warmed off-white. The AI default reaches for blue; the good sites increasingly reach for warmth. The AI default sets headlines medium-bold; the good sites set them light and large with negative tracking. The AI default mixes radii and scatters the accent; the good sites enforce one of each. The generic look is not mysterious. It is a list of unmade decisions.
You do not escape the average by switching tools. You escape it by removing the ambiguity the model fills with defaults. Give it values instead of adjectives: exact background hex, one accent used only on actions, a named font with size, weight and tracking, one radius, a real motion budget — and a "never" list that closes the exits (no gradients on white, no emoji icons, no three identical cards). I wrote the practical version in how to avoid the generic AI look.
The catch is that most people cannot produce those values from memory. That is the entire reason Norrly exists: to measure the sites you already admire and hand you their decisions as a paste-ready spec. Point the model at a designed position — a real palette, a real type scale, a real rhythm — and it stops giving you the median and starts giving you something with a point of view. The tool was never the problem. The blank was.