uptick systems · release notes · experiment 001
A landing page that runs its own growth experiments.
Shipped August 22, 2026 · live at uptick.systems/how
We built a page that helps anyone figure out how to actually use AI, and then we gave the page itself an AI research team. It writes its own copy variations, tests them on real visitors, kills the losers, and keeps a journal of what it learned. Every night. No humans in the loop.
It exists to test one thesis: most people do not know what AI can do for them, so the barrier is not price or trust, it is orientation. This page removes every hurdle between “vaguely curious” and “talking to an AI that shows me.”
You land, you tap one button, and you are in a live voice conversation (typing works too, no mic required). A friendly guide asks about your week, what you keep putting off, where the time goes. It never asks for a goal, an email, or any setup.
As it spots opportunities, it saves them as recipe cards: a ready-to-paste prompt written around your actual life, plus what to expect when you use it. The conversation ends with your recipes on screen, copyable, and an optional “email me these” box. That is the whole funnel. No form gates, no links out, no pitch.
This is the part worth showing off. The page is a closed-loop optimization system with an explicit objective:
Four hand-written page variants went live at launch, spanning different headlines, button copy, conversation openers, and three different page layouts. From there the machine takes over:
A Thompson-sampling bandit picks which variant to show, favoring variants that are earning more reward while still giving new ones a fair shot.
Impressions, conversation starts, recaps, and emails are recorded per variant, deduplicated per visitor session.
An AI researcher reads the score table, retires any variant that is statistically dominated (200+ views and less than 5% chance of beating the leader), and writes new challenger variants to replace them.
Every new variant passes a hard-coded guardrail check before going live: no prices, no hype, no fake urgency, and it can never violate the core rule that visitors are never asked for goals or preparation. Every action is logged to a research journal we can audit.
So the site gets measurably better at its job while we sleep, and we can read its reasoning in the morning.
| Can mutate | Can never touch |
|---|---|
| Headline, subhead, button copy, microcopy | The reward function and success metrics |
| The guide's opening tone and first question | Cost caps and rate limits |
| Page layout, from three pre-built structures | The no-goals, no-prep conversation rules |
| Anything outside its guardrailed word budget |
Live in production since August 22. Seed variants serving, events recording, first autonomous research run scheduled nightly. Voice sessions are capped at ten minutes with a graceful wrap-up, and the whole experiment has hard rate and cost limits.
try it
uptick.systems/how — the experiment (talk to it, honestly, it is the best way to get it)
uptick.systems/discover — the browse-first sibling it is competing against