The short answer
A language model predicts the most likely output. Trained on the whole web, the "most likely" landing page is the average one - so when you ask for "a clean, modern site", it hands back the pattern it has seen ten thousand times. That average has a recognizable shape, and once you notice it you see it everywhere.
The tells of the AI default
- A centered hero with one big headline and a subhead.
- An "eyebrow" pill or badge above the headline.
- A gradient accent word in the middle of the title.
- Three feature cards with small icons.
- Pill-shaped buttons and rounded-everything, on a near-white or near-black ground.
None of these are wrong on their own. The problem is the combination is a default, so a site built entirely from defaults reads as generic no matter how polished it is.
Why "make it look better" does not fix it
Adjectives like "modern", "clean" or "premium" don't move the model off the average, because the average already looks modern and clean. To get somewhere specific you have to give the model something specific: real colors, real fonts, a real layout to work from.
How to break out of it
- Pick a site whose look you actually like.
- Pull its real style with the style extractor - exact colors, fonts, type scale and tokens.
- Paste that into your AI along with an instruction to apply the direction to your own content, and to avoid the default patterns above.
- For repeatable results across projects, capture references into taste profiles with the extension and feed them to your AI over MCP.
The fix is grounding. Once the model is building from concrete references instead of the mean, the output stops looking like everyone else's.