Can taste be turned into data?

Ask an AI tool for a website and you'll get one fast. It'll be tidy, technically fine and a lot like every other site it has ever made. Making things is easy for machines now. Telling the good from the merely average is where they fall down.

Taste Labs is betting that's a solvable problem. The San Francisco company calls itself "the taste layer for AI," and its mission line doesn't hedge: "Ending AI slop." It launched last October, raised an $18.5 million seed round co-led by CRV, and, according to a Business Insider feature this week, already brings in "tens of millions in revenue."

How you teach a model taste

The founder and CEO, Thais Castello Branco, previously worked at Kraft Heinz and was an early employee at the AI search company Exa. Her thesis is that taste is a form of pattern recognition, so it can be broken down and learned.

The way in is people. Taste Labs works with a network of about 1,000 Tastemakers, contractors from creative fields who design things and explain their choices, grade AI outputs and review each other's grades. An art director who has worked for Adobe and Adidas told Business Insider the work means breaking design into observable parts like typography, motion, color and texture, and that "it's not as simple as just rating out of 10."

All of that becomes training data. The company says it works with "the top frontier labs," though it won't say which, starting with product websites, slide decks and brochures before moving into copywriting, fiction, music and film.

The Brand API

The part you can actually use today is the Brand API, currently in beta. It does three things:

• Brand extraction: give it a URL and it returns the whole brand system as structured data: logos, color palettes, fonts, motion, textures, shadows and more, plus screenshots and the captured HTML and CSS.
• Style search: describe a look in plain words, like "a dreamy brand" or "dark brutalist developer tools," and it finds matching brands in a curated index.
• Verify adherence: compare a new page against a reference brand and get a score, a written critique and a to-do list of fixes, worst first.

It's built for developers and AI agents. There's an MCP server that works with Claude Code, Cursor and Codex, so an agent can pull a brand before it builds and check its own work after.

Pricing

You get 20 free credits to try it. After that, Starter is $99 a month for 600 credits and Pro is $749 a month for 6,000. An extraction or a verification costs 2 credits, and a search costs 1 or 2 depending on depth.

Heads up

This is infrastructure for people building AI tools. Unless you're building with AI agents, the Brand API is more interesting to read about than to use, and two of its three tools are still marked alpha.

The bigger question is the one Business Insider raised: do we want AI to define what's great? Castello Branco's answer is that the goal is raising the floor, getting everyday AI output to a baseline of quality, and helping people who can't code or draw get closer to what they had in mind. That's a reasonable aim. It's also worth noticing that taste usually comes from years of looking, making and getting it wrong, and no amount of graded data does that part for you. (More on that in When you can make anything, the job is knowing what's yours.)

One finding from the company's own research is worth keeping in mind: after analyzing more than two million websites going back a decade, Taste Labs says the web was already becoming more alike before AI arrived.

Who it's for

Developers and teams building AI tools that make visual work, who want those tools to stay on brand. Design leads curious about where AI design is heading. And designers who'd like to be part of it: the Tastemakers network takes applications, and it pays people for exactly the judgement this whole industry runs on.

If taste is the thing you care about, this is a company worth watching closely, and arguing with.

Written by Steve Salgado