An AI partner for better experiments.
Darwinch is being built as an AI experimentation partner that works alongside agents in your environment, helps choose useful tests, and learns from the evidence over time. The aim is to make sound experimentation easier without asking you to become an analytics expert.
Available in the pilot
The current pilot supports AI-assisted headline and call-to-action text ideas. You preview the exact change, approve it, and choose when to start the experiment.
Start with a hypothesis and one browser-reported click goal. Keep the original text as the control. AI can suggest up to two plain-text alternatives, or you can write your own. Approval and starting are separate actions; AI cannot publish.
Delivery changes one marked heading, link, or button. Missing targets, incompatible markup, and blocked or failed delivery leave your original page usable. You can pause the experiment and restore the original.
Evidence you can use
The pilot compares one control and one approved variation at a fixed 50/50 allocation, with a 14-day enrollment window, a 24-hour conversion window, and at least 2,000 assignments per version before a conclusion.
Assignments, successful exposures, and deduplicated conversions are reported separately. Failed delivery stays in the assignment denominator. Results remain preliminary until the planned stopping rule, and delivery or sample problems can block a conclusion.
Results measure browser-reported clicks, show uncertainty, and can remain inconclusive. They do not establish verified purchases, leads, or human visitors. No winner is promoted automatically.
The direction from here
Agent integration, supervised iteration, and an autopilot mode are planned. They are not available in the current pilot.
The intended path is an agent that can help form hypotheses, suggest the next useful test, and refine recommendations as evidence arrives. Any future autonomy needs explicit boundaries for approval, spending, measurement, and rollback.
Permission comes first
The browser SDK starts with measurement off. Consent, Global Privacy Control, Do Not Track, and opt-out settings govern delivery. Darwinch works independently of Glowcrumb; session replay is not required.