The Hypothesis

Can a landing page rebuild itself autonomously using behavioural data — without changing its URL or disrupting ad campaigns?

The Concept

Traditional A/B testing for landing pages is manual, slow, and disconnected from the ad campaigns driving traffic. This experiment builds an autonomous system where a single stable URL (e.g. `domain.com/landing/offer01`) serves different page variants via server-side routing. The system captures granular behavioural data — clicks, scroll depth, rage clicks, hover duration, text highlights — and an AI agent uses that data to generate challenger variants with modified copy, layout, CTA placement, and visual hierarchy. Champions are promoted automatically. The ad campaign never knows the page changed.

Self-optimising landing pages

The hypothesis

Can a landing page rebuild itself autonomously using behavioural data — without changing its URL or disrupting ad campaigns?

The concept

Traditional A/B testing for landing pages is manual, slow, and disconnected from the ad campaigns driving traffic. This experiment builds an autonomous system where a single stable URL (e.g. domain.com/landing/offer01) serves different page variants via server-side routing. The system captures granular behavioural data — clicks, scroll depth, rage clicks, hover duration, text highlights — and an AI agent uses that data to generate challenger variants with modified copy, layout, CTA placement, and visual hierarchy. Champions are promoted automatically. The ad campaign never knows the page changed.

How it works

1. Stable URL receives traffic — same endpoint, ad campaigns untouched
2. Capture behavioural data — clicks, scroll, dwell time, rage clicks, highlights
3. AI analyses friction points — where users drop off, hesitate, re-read
4. Generate challenger variant — modified copy, layout, CTAs, social proof placement
5. Split traffic server-side — champion vs. challenger behind the same URL
6. Promote or discard — statistical significance → new champion or kill

The loop runs continuously — pages literally rebuild themselves guided by real user behaviour.

What it explores

• Can autonomous page generation match or beat human-designed landing pages?
• How much behavioural signal (sessions, time) is needed before a challenger is viable?
• Does server-side variant routing maintain ad platform attribution integrity?
• What’s the right balance between structural changes and copy-only changes?
• Can institutional memory from past pages guide new page generation?

What we found

• 14 autonomous rebuild cycles completed across 3 campaigns over 60 days
• Conversion rates improved 41% from the original human-designed pages
• Biggest gains came from CTA repositioning and social proof placement — not copy changes
• Server-side routing preserved ad attribution perfectly — no disruption to Google Ads or Meta campaigns

Learnings

• Behavioural data (scroll depth, rage clicks) is a stronger optimisation signal than click-through alone — invest in granular event capture.

• Budget ~800 sessions per variant before trusting statistical significance — below that threshold, the system makes confident but wrong decisions.

• Lock brand CSS as an immutable constraint — it prevents the AI from drifting off-brand during autonomous rebuilds without limiting structural experimentation.

• A shared asset library with AI-tagged metadata is essential — the agent needs to know what images are available and what they represent.

• Institutional memory across pages compounds — persist learnings from campaign N so they improve the starting point for campaign N+1.

Where it goes next

This is now a core feature of Flywheel. The open thread: can the same autonomous rebuild loop work for email templates and ad creative — not just landing pages? The behavioural data is different (no scroll depth in email), so the signal model needs rethinking.