← Back to Field Notes
It started with a frustration.
We were watching a client’s marketing team run A/B tests on a landing page. Two variants. One had a different headline. They split traffic 50/50, waited two weeks, picked a winner. Then they started the next test — this time the CTA button colour. Another two weeks.
At that pace, they’d test maybe 20 variations a year. On one page. And most of the “variations” were cosmetic — headline swaps, button colours, hero images. Nobody was testing structural changes because those take a designer, a developer, and a sprint cycle.
We started wondering: what if the page could just do this itself?
Not A/B test in the traditional sense. Actually rebuild. Look at what’s working, look at where people are dropping off, and generate a new version — layout, copy, CTA placement, everything. Behind the same URL, so the ad campaigns never know the difference.
We spent two weeks building a prototype. Stable URL, server-side routing, behavioural capture (scroll depth, rage clicks, dwell time), and an AI agent that generates challenger variants based on friction signals.
After 60 days across three campaigns, the system completed 14 autonomous rebuild cycles. Conversion improved 41% from the original human-designed pages. The biggest gains came from CTA repositioning and social proof placement — not the copy changes we expected.
The full experiment is in the sandbox: Self-Optimising Landing Pages. But the thing we keep coming back to is this: the system didn’t find one winning page. It found that the concept of a “final” landing page is wrong. The best version is always the next version. The page is never done — it’s just current.
That idea is now core to how Flywheel thinks about every asset it generates.