Sample efficiency
Built for the traffic you actually have.
Classic experimentation needs tens of thousands of visitors before it can tell you anything. Most surfaces never see that. Curious is engineered, end to end, to learn from the traffic that shows up.
The problem
Why A/B testing fails below scale.
An A/B test has no answer until it reaches sample size. At a 5% conversion rate, detecting a 10% relative lift takes tens of thousands of visitors per variant. On a surface that sees a few hundred visitors a day, that's months for a single test, and most tests lose. So teams below that scale conclude experimentation isn't for them, and every decision stays a guess.
The math is the problem. So Curious changes the math.
How
Six ways Curious changes the math.
Starts informed.
Before your first visitor arrives, Curious builds a statistical prior from your business context: who your users are, what typically converts for surfaces like yours, and what each variant is likely to do. The first day of traffic refines an informed starting point instead of feeding a blank slate.
Wastes less traffic.
An A/B test spends half its traffic on the loser by design. A contextual bandit shifts traffic toward what performs as evidence accumulates, so fewer visitors are spent proving what's already becoming clear.
More signal per visitor.
Curious removes predictable user-to-user variation before estimating what a variant did (CUPED-style variance reduction). The same visitors carry more information, so conclusions arrive sooner.
Learns from early signals.
A purchase takes days to observe and churn takes months. The clicks, sessions, and trial starts that precede them arrive in minutes. A surrogate model learns from those early signals now and is checked against the real outcome when it lands.
Every metric on its own clock.
A one-day click signal doesn't wait for a thirty-day refund window to close. Each reward matures independently, so fast evidence flows into the model while slow evidence is still arriving.
Opinion is capped. Data isn't.
Priors and predictions are engineered so model judgment contributes bounded evidence. Only real outcomes accumulate without limit. AI can speed up learning, but it can never outvote your users.
The honest part
None of this repeals statistics.
With ten visitors a day, learning is still slow. Confidence still comes from real outcomes, and Curious never declares a winner your data doesn't support. What changes is where learning starts, how much each visitor contributes, and how little is wasted along the way. On low-traffic surfaces, that's the difference between experiments that conclude and experiments that would never have been run.