We thought we'd build AI on top of what we already had. The benchmark told us otherwise.

Act 1 — Trying to force AI into a block editor

We started where it was natural to start. Inside the page builders we'd already been shipping. SeedProd's own block-based builder. Block markup, broadly. Two things kept breaking. Block syntax is verbose, so the model burned tokens producing valid markup and still got it wrong. And even when it worked, the designs felt boxed in. We were putting AI in a cage.

Act 2 — The benchmark

So we ran a benchmark across every site builder we could find: block-based WordPress builders, Wix, Squarespace, Webflow, plain HTML and Tailwind, several static-site frameworks. Plain HTML with Tailwind didn't just win. It won by a wide margin.

Why? Two reasons. AI has years of training data on plain HTML and Tailwind utilities. It's already fluent. And AI loves a file system. Read a file, edit a file, run a command, take a screenshot to verify the result. That's how engineers build software, and it's how AI builds best.

Act 3 — Building SiteShip

We stopped trying to adapt AI to a website builder. We built an agent-native website engine that adapts to AI. We rebuilt the foundation around the parts AI is best at, and gave the agent the same kind of toolkit a developer uses, plus a library of tested patterns to follow on every job.

The result: AI rarely fails. Sites come out looking great the first time. The same conversation that builds the site keeps editing it forever.