The Landing Page That Optimizes Itself: An Autonomous Growth Experiment
Corrado Giaquinto, The Birth of the Sun and the Triumph of Bacchus, 1761
Vaughn DiMarco
We gave a landing page an explicit reward function and an autonomous research loop: a bandit serves competing page variants to real visitors, and every night an AI researcher retires statistically dominated variants and writes guardrail-checked replacements, logging every decision to an auditable journal. The page improves against its objective while we sleep.
The experiment: a page with its own growth team
This week we shipped uptick.systems/how, a page that helps anyone figure out what AI can actually do for them: you tap one button and you are in a live voice conversation with a friendly guide that turns your week into copy-paste AI recipes. No form, no email gate, no prep. That alone is a bet worth testing. But the more interesting part is underneath: the page runs its own growth experiments, end to end, without us.
Every visitor is served one of several competing versions of the page. Every meaningful action is scored. And every night, an AI researcher reads the results, kills the versions that are statistically losing, writes new challengers, and records its reasoning in a journal we can audit over coffee. We wrote the first four variants by hand on launch day. Everything after that is the machine iterating against a number we chose.
Start with an explicit reward function
Most conversion optimization fails before it starts because nobody wrote down what a conversion is worth. We gave the page one objective, in one line: reward per visitor equals 1.0 for starting a conversation, plus 0.5 for staying to the recap, plus 1.0 for leaving an email. Starting the conversation dominates on purpose, because the whole thesis of the page is that the barrier to AI adoption is orientation, not price, and orientation begins the moment someone actually talks to the thing.
Writing the reward down does two jobs. It makes every downstream decision mechanical: the traffic allocator and the researcher optimize the same number, so there is no room for the quiet goal-drift that turns growth experiments into vibes. And it makes the system honest. A page could juice conversation starts with clickbait, but starts that never reach the recap earn less than half the available reward, so shallow tricks lose on the scoreboard they are judged by.
A bandit serves the variants, not a calendar
Classic A/B testing splits traffic evenly for a fixed period, then a human reads a dashboard and picks a winner. We used Thompson sampling instead: each variant’s performance is a probability distribution, and every visit the system samples from each distribution and serves whichever variant sampled highest. Strong performers earn traffic automatically; weak ones fade without a meeting. New variants start with an optimistic prior, so fresh ideas get a real audition instead of being starved by the incumbent.
The variants are not just headline swaps. Each one is a bundle of copy, the voice guide’s opening tone, and one of three page structures: a talk-first layout, a type-first layout for the microphone-shy, and a layout that shows sample recipe cards before asking anything. The system is allowed to discover that the winning page is a combination no one on the team would have picked.
The nightly researcher, and the cage it works in
Once a day, a language model gets the score table and a mandate: analyze what is winning and why, retire any variant with at least two hundred views and less than a five percent chance of beating the leader, and if the pool is thin, propose replacements. Its proposals are hypotheses grounded in the data it just read, and its analysis is saved to a research journal, so the reasoning behind every change to the page is inspectable after the fact.
Autonomy without limits is how you wake up to a page you are ashamed of, so the researcher works inside a cage it cannot reason its way out of. Every proposal passes a hard-coded validator before going live: no prices, no fake urgency, no hype vocabulary, length caps on every field, layouts only from the pre-built set. And it can never touch the rules that define the product, like the one that says the guide never asks visitors for a goal or implies they should have prepared. The reward function, the cost caps, and the rate limits live outside its reach entirely. The model proposes; the code disposes.
Why this pattern matters beyond one page
Strip away the specifics and the architecture is: an explicit reward function, a bandit allocating real traffic, an AI researcher iterating nightly inside guardrails, and an audit journal. Nothing in that list is specific to a consumer landing page. Swap the reward for booked discovery calls, qualified signups, or completed onboarding, and the same loop optimizes a B2B funnel, a pricing page, or an activation flow. The expensive part of conversion work was never the ideas; it was the coordination loop of noticing, deciding, writing, shipping, and measuring. That loop is now a cron job.
This is AI BizOps applied to growth: an agent owns an operational workflow end to end, a human owns the objective and the guardrails, and the output is measured in a number the business already cares about. The interesting question stops being whether AI can write better landing page copy than you. It is how many of your funnels are still waiting for a human to find time to iterate on them at all.
What we are watching, and how to follow along
The experiment is young: the seed variants are serving, events are recording, and the researcher makes its first unsupervised call soon. We are watching three things. Whether conversation-start rates differ meaningfully across structures, which would validate that layout, not just copy, drives orientation. Whether the researcher’s generated variants can beat hand-written seeds, which is the real test of autonomous iteration. And whether the talk-to-it page outperforms its browse-first sibling at uptick.systems/discover, two philosophies for the same audience running head to head.
The page is live at uptick.systems/how, and the full technical writeup, including the reward math and exactly what the researcher can and cannot touch, is at uptick.systems/how/notes. Try the conversation itself; it remains the fastest way to understand what we are testing. And if you are wondering what this pattern would do pointed at your own funnel, that is exactly the kind of question the instant AI audit exists to answer.
Common questions
How is a bandit different from an A/B test?
An A/B test splits traffic evenly for a fixed period, then a person reads a dashboard and picks a winner. Thompson sampling treats each variant as a probability distribution and shifts traffic toward whichever samples highest on each visit, so strong variants earn traffic without a meeting.
What stops the AI researcher from shipping bad copy?
Guardrails it cannot edit, plus an audit journal. It can only retire variants that are statistically dominated and write replacements inside fixed limits, and every decision it makes is logged with its reasoning so a human can read it back.
Does this pattern only work on landing pages?
No. The parts are a reward function, a bandit, a nightly researcher, and an audit journal. Swap the reward for booked calls, qualified signups, or completed onboarding and the same loop runs against a B2B funnel, a pricing page, or an activation flow.
Find out what agentic workflows would save your team.
The $999 assessment identifies 5-10 hours per week of recoverable time, with a money-back guarantee if it doesn't.
Get the Assessment →