Every website builder will happily generate you a page. The interesting question is what happens on day two, when the page is live and you’re wondering if the headline is costing you money.
Most tools answer with a report. As of today, a ONE page answers by fixing itself.
One prop under test
Open your page in the editor, hit Experiments, and pick one thing — the headline, the button label, the price line. Write a few variants, or leave the box empty and let the AI propose some. That’s the whole setup.
From that moment:
AI proposes variants (offline)
│
▼
visitors split by a hash
same visitor, same version, every time
│
▼
every conversion tagged with
the version that produced it
│
▼
hourly: is any variant a
statistically significant winner?
yes → promoted into the page
no → keep testing
significant loser → retired
The AI appears exactly once in that diagram, and it’s offline. It drafts candidate copy into a pool before any visitor arrives. The live path — which version you see, how it renders, how the click is counted — is deterministic. No model in the hot path, no flicker, no third-party script.
The ratchet
The promotion rule is one function, and it’s deliberately strict: a variant replaces the original only when its lift is statistically significant and positive. Not “trending up.” Not “looking good after lunch.” Significant.
If nothing clears the bar, the experiment just keeps running. If the whole pool turns out to be significantly worse than what you had, it’s dissolved and your original stands. And when a winner is promoted, the value it replaced — and the measured lift — are written into the experiment’s history. The page can only move up, and you can always see the receipts.
We call this shape a ratchet, and it’s the same discipline we hold our own build system to: change is cheap, regression is forbidden, and every improvement is a number before it’s a decision.
Why code generators can’t follow
Tools like Lovable and v0 generate code — hundreds of lines of JSX per page. That’s fine until you want a bot to improve the page in production, at which point you’d be letting a language model rewrite live source code unsupervised. Nobody sane does that.
A ONE page isn’t code. It’s a small, typed tree of data — component names and their props — rendered by a fixed library. Changing the headline is changing one field in one record. That’s an operation a bot can do safely, reversibly, and under a statistical gate.
The page stops being a deliverable you received once, and becomes an asset that improves while you’re doing something else.
What you’ll see
In the editor, the Experiments panel shows each arm with its sessions, conversion rate, lift, and a significance badge — plus the state: testing, won, or dissolved. On a won experiment you’ll see exactly what was promoted, when, and at what measured lift.
That’s the whole feature. One prop, one real signal, one promotion rule that refuses to guess. Point it at your busiest page and let the math argue with your headline instead of your team.