The taste problem

Why AI-built websites look generic

AI website builders look generic because they have no reference for craft: with nothing specific to ground them, they fall back on the most common shapes in their training data, the same centred hero, purple gradient, Inter, and three identical cards on every brief. A taste layer fixes it by feeding the model 268 deeply-analysed real sites, so it composes from how the best sites are actually built instead of from the average.

Generic AI build defaults

Build Better Today

The all-in-one platform to grow your business faster than ever.

Get Started

Feature One

Lorem ipsum dolor sit amet consectetur.

Feature Two

Lorem ipsum dolor sit amet consectetur.

Feature Three

Lorem ipsum dolor sit amet consectetur.

Grounded by Palate taste
Studio

We grow the businesses
others overlook.

A practice built on attention, not templates. Three decades of compounding craft.

See the work Read the approach

01

A considered, asymmetric system

Same brief. The left is the default an AI builder reaches for: a centred hero on a purple gradient, Inter, three identical cards. The right is grounded in Palate's references: an editorial type system, an off-centre composition, one disciplined accent, and real inner-page depth.

The default is an average

A model with no grounding reproduces the centre of its distribution. For websites that centre is a hero over a gradient, a system sans, a three-card feature row, and a section order it has seen ten thousand times. It is competent and forgettable, the visual equivalent of "leverage" and "seamless".

Taste is specific, not abstract

Craft does not transfer as advice ("make it modern"); it transfers as specifics: this type scale, this off-centre composition, this one disciplined accent, this inner-page depth. Palate captures those specifics from real sites as structured data an AI can actually compose from.

What's in the library

References268
Verticals25
Page types27
Flagship-tier69

Each reference carries design tokens, do and don't rules, signature moves and section anatomy. Browse them in the collection.

How a taste layer works

Palate is a hosted MCP (Model Context Protocol) server. Your AI builder calls it for a brief, gets a backbone plus diverse donors with explicit borrow rules, composes from at least three of them, and re-skins every identity layer. The output is grounded and one-of-a-kind, not averaged.