How We Built an AI That Edits Your Storefront in Real Time

When we started designing the AI Website Builder, the hardest problem wasn't getting an AI model to generate good HTML. It was getting merchants to trust it enough to actually click publish.

Most AI page builders treat generation as a one-shot event: you type a prompt, wait, and get a full page back. If it's wrong, you start over. That works for a demo. It falls apart the moment someone is editing a storefront that's actively taking payments.

Editing, not generating

So we reframed the problem. Instead of asking the model to regenerate a whole page, every prompt is scoped to a diff against your existing storefront: add this section, change this color, reorder these plans. The model only ever proposes a change, never a rewrite.

• Every request is parsed into a small, reviewable diff
• Changes render in a live preview before anything touches production
• Every change is versioned, so undo is a single click, not a support ticket

The AI doesn't need to be right on the first try. It needs to be safe to try.

What's next

We're currently testing multi-step prompts, where a single request like "redesign my pricing section to highlight the annual plan" gets broken into a sequence of smaller diffs you can approve individually. Expect that in a future release.

More from the blog