TL;DR
- Real difference comes from changing the model's constraints — paste in a design system and the convergence breaks
- AI-generated sites converge on the same look because the model averages its training data — the convergence is mathematical, not aesthetic
- Adding adjectives to prompts ("unique", "premium", "different") produces slightly-different averages, not actual difference
If you have built more than one site with Lovable, v0, or Bolt, you have probably noticed that they all look like cousins. Different colors, different copy, but the same underlying shape. Same hero rhythm. Same card style. Same button radius. Same Inter typography. There is a reason for this and it is not laziness.
The math behind the sameness
LLMs are trained to produce the most likely continuation of a prompt. "The most likely landing page" is, statistically, the average of every landing page in the training set. That average is heavily skewed toward a small number of high-influence reference points — Stripe, Linear, Vercel, Tailwind UI templates, and the long tail of YC sites that imitate them.
When you ask Lovable to "build a SaaS landing page," it does not pick a single reference and copy it. It interpolates between thousands of references and produces the centroid. The centroid is the AI-built look. Every Lovable user who runs the same kind of prompt gets a sample from roughly the same distribution, which is why the sites all look like cousins.
Why adjectives do not work
The first instinct is to fight the averaging with adjectives. "Make it unique." "Make it premium." "Make it different from a typical SaaS site." These do something — but not the thing you want. They shift the model to a slightly different averaged region (premium = darker palette + serif headings + more whitespace), but the output is still an average. It just looks like a more expensive version of the same site.
Real difference comes from changing the model's constraints, not its adjectives. The model is producing the most likely page given the constraints it has. If the only constraints are "SaaS landing page" plus your copy, the most likely page is the centroid. If you add real constraints — specific tokens, specific component shapes, specific layout grammar — the most likely page given *those* constraints is no longer the centroid.
What real constraints look like
A real constraint is something the model would not have inferred on its own. Examples:
rounded-2xl everywhere but a specific radius scale tied to component size.You can write all of these by hand. You can also paste them in as a pre-built pack. The Rottoways design system pack is the latter — a coherent set of constraints designed to push AI builders out of the centroid and into something that looks intentional. The same logic applies to motion: UI Motion Prompts supplies the motion-layer constraints that Lovable and Cursor would not have inferred.
The test
After you have added real constraints, the test is simple: open your site next to five other Lovable sites and ask whether yours stands out. If it does, the constraints worked. If it still looks like a cousin, you have not changed the visual vocabulary enough yet — most likely you still have default Inter, default rounded-2xl, or the default three-column feature grid.
Stop sampling from the average
AI builders are extraordinary at speed and adequate at design. The path forward is not to abandon them or to "design properly" before using them. It is to give them better constraints. Two paste-ins handle most of the work.
Rottoways supplies the design constraints. UI Motion Prompts supplies the motion constraints. Together they pull your AI-built site out of the average and into something that looks like itself.
More articles
Why Every Lovable Landing Page Looks the Same
Lovable defaults produce a recognizable visual fingerprint — the same gradient hero, the same pill buttons, the same boxy card grid. Here's why, and how to break the pattern.
Best Vibecoding Tips to Improve Your AI-Generated Website
A practical checklist of vibecoding upgrades — design system, motion, copy, and layout — that turn a generic AI-generated site into something that looks like it was built by a human.