Absolutely Butter

Know which version is winning before you have the traffic to prove it.

A/B testing tools built for enterprise scale need thousands of conversions before they’ll say anything. Absolutely Butter gives you an honest probability from the traffic you actually have.

Start free trial

Drop in the SDK, ship the test

Three calls. No provider component, no config file, no build step. The SDK has zero dependencies and always falls back to your control, so a bad network never breaks the page.

app/signup-cta.tsx
import { init, getVariant, track } from '@absolutely-butter/sdk'

// Both values below are placeholders — copy the real ones from your dashboard.
await init({
  apiKey: 'pk_live_xxxxxxxxxxxxxxxxxxxxxxxx', // Settings → API key
  experimentId: 'exp_xxxxxxxxxxxx',           // the experiment's detail page
  baseUrl: 'https://api.absolutely-butter.com',
})

// 'control' | 'variant' — synchronous, never throws, never null
const variant = getVariant()

// call this when the visitor completes the goal you're testing
track('conversion')

Statistics that stay honest at low traffic

Absolutely Butter uses Thompson sampling and a Beta–Binomial model. In plain terms:

  1. 01

    Standard tools assume you wait, then look once

    Classic A/B testing math is only valid if you fix a large sample size up front and check the result a single time at the end. Founders check every morning on a few hundred visitors — and every extra look inflates the odds of a false winner.

  2. 02

    We model each variant as a probability, updated every visit

    Instead of a pass/fail verdict, each variant carries a distribution over its true conversion rate. Every impression and conversion sharpens it. Looking early changes nothing, because nothing is waiting to be "triggered".

  3. 03

    “78% probability variant is better” means exactly that

    Of all the true conversion rates still consistent with your data, the variant beats the control in 78% of them. It is a direct statement about your experiment — not a p-value, not a significance threshold you either cleared or didn’t.

  4. 04

    Wide ranges early on are honesty, not a bug

    At low traffic the credible intervals we show are wide and often overlap. That is the tool telling you it does not know yet. They tighten on their own as data arrives — a narrow range you hadn’t earned would be the real defect.

The full derivation — expected loss, the decision label, the forward projection — is in the docs.

One plan

$19/ month

30-day free trial. No credit card required.

  • Unlimited experiments and traffic
  • The full stats engine — probabilities, expected loss, credible intervals, forecasts
  • The zero-dependency SDK and the dashboard
  • Automatic traffic allocation via Thompson sampling
Start free trial