Why AI-built websites all look the same
TL;DR
AI-built sites share a look because, with no direction, a model converges on the average of everything it has seen, and the average is a template. The model is not the limit: the same assistant produces a distinctive site when you ask for one strong compositional move, real images at scale, and copy a competitor could not reuse.
The tells
Once you have seen a few, you recognise them instantly. Not because they are ugly, but because they are all making the same safe choices.
- A centered hero with a headline, a subheadline and two buttons
- Three identical cards directly underneath, each with a generic icon
- Rounded corners and a soft shadow on absolutely everything
- A purple or blue gradient somewhere, doing no particular job
- Paragraphs running the full width of the screen
- A big title with default letter-spacing, which is the quietest tell of all
- Stock imagery, or no images at all
- Copy about excellence, passion and tailor-made solutions
Why it happens
A model asked for "a nice website" does the statistically safe thing: it produces the average of everything it has seen. The average of every landing page on the internet is, almost by definition, a template. Nothing went wrong. The absence of a direction is itself a direction, and it points at the middle.
The same mechanism explains the copy. With no facts about your business, the safest sentence is the emptiest one, so you get "your satisfaction is our priority" rather than a number that only you could write.
It is not the model, it is the direction
This is worth being clear about, because there is a whole industry built on the opposite claim. Nobody has a smarter model than the one you already pay for. The difference between a generic result and a distinctive one comes from what the model is asked to do, and from whether anyone looks at the output before it ships.
Same assistant, different instructions, unrecognisably different site. That gap is the entire craft.
What to ask for instead
- One strong compositional move in the hero: a giant wordmark, a full-bleed photograph with the title biting into it, an asymmetric split, or a single huge figure. Left-aligned to the gutter, not centered.
- An editorial list instead of a card grid: full-width rows, title left, description in the middle, hairline separators. It reads as designed rather than assembled.
- Real images, at scale, worked into the layout. One photograph of your actual work does more than any amount of spacing tuning.
- Copy with numbers and names in it, and nothing a competitor could paste onto their own site unchanged.
- A line-length limit on paragraphs, and controlled tracking on large titles.
- One section that is almost empty. Restraint is the hardest thing to get from a model, and the most convincing when you do.
Then make it look at its own work
The last step is the one almost nobody takes. Ask the assistant to stop building, change role, and review what it produced as an art director would, scoring it honestly and fixing the weakest point before publishing. Two passes is usually the difference between a first draft and something you would put your name on.
The full grid is in the guide on reviewing before you publish, and the rest of the sequence, from the brief you give first to making the site findable, is in the six guides.
Questions fréquentes
Why do AI-generated websites all look alike?
Because a model given no direction produces the average of everything it has seen, and that average is a template: centered hero, three cards, generic icons, soft shadows. The output is safe rather than wrong, and safe looks identical every time.
Can AI make a website that does not look AI-generated?
Yes, with the same model. Ask for one strong compositional move, an editorial layout instead of a card grid, real images at scale, and copy carrying numbers and names. Then have the assistant review its own output before publishing.
Does a better AI model fix it?
Not really. The gap between a generic and a distinctive result comes from the instructions and from reviewing the output, not from the model. The assistant you already pay for is capable of both.