ProveBetter

A/B testing that tells you what wins

Two statistical engines. One plain-language answer.

Optimizely shows p-values; VWO shows chance-to-beat — nobody shows both. ProveBetter runs the frequentist and Bayesian engines side by side on every test, each verdict in plain language. Decide on evidence, not an interpreter.

Free tier at launch · No credit card · A first test takes minutes, not a procurement cycle

How it works

  1. 1

    Create a test

    Name it, pick 2–5 variants, split the traffic.

    Open the builder →
  2. 2

    Paste one snippet

    Your variants ship inside the generated block — done in minutes.

    See the snippet →
  3. 3

    Read the plain-language winner

    Two stat engines agree on the answer, in words.

    Browse tests →

Want to see it with data first? Explore the demo experiment →

No flicker.

The boot script is about 1.4 KB gzipped and hides test elements for at most 2 seconds while your visitor's variant loads — on a slow connection visitors simply see the control page and are never counted. Site speed stays part of the experiment, not its casualty.

No engineers required.

Client-side tests take one snippet: write your variants in the builder, paste the generated block once, edit copy in the dashboard from then on. Server-side tests take a few lines of API code. No demo, no implementation project.

No sales call.

Pricing is published on our pricing page — priced on events, not tracked users, so your bill grows with the value you measure. Free tier available today, no credit card. No quote-only contracts, no “contact us” form standing between you and a test.

Everything a trustworthy test needs

Shipped in the product today — not a roadmap slide.

2–5 variant tests

Run A/B, A/B/C, up to A–E from day one. Multi-arm isn't an enterprise add-on.

Client + server modes

One builder for web experiments and server-side feature tests. No split products, no migration project.

Live traffic dial

Set each variant's traffic percentage at creation and adjust it mid-test — changes apply instantly.

Sticky assignment

SHA-256 hashing plus a first-party cookie keep every visitor in their variant across visits and traffic changes.

Dual stats engines

Frequentist z-test (p-value, confidence interval) and Bayesian posteriors (credible interval, chance-to-beat, expected loss) on the same screen, every verdict in plain words.

Scheduling

Set a start and end date up front; the test begins and ends on schedule, and out-of-window visitors are never counted.

Winner rollout in one click

Declare a winner and it rolls to 100% of in-test traffic — or clear it and keep testing.

SRM + correction guardrails

Automatic sample-ratio-mismatch checks flag broken assignment, and multi-variant p-values get multiple-comparison correction before you ship a bad result.

Frequently asked questions

Is it really free to start?
Yes — free tier available today, no credit card, and the full prices are published on the pricing page: no “contact sales”, no quote-only contracts. We price on events, not tracked users — for contrast, ABsmartly starts at €60K/yr with no trial, Optimizely is custom-quoted from around $36K, and VWO prices per tracked user.
How is this different from Optimizely, VWO, and ABsmartly?
Three concrete differences. Optimizely splits web and feature testing into separate products and won't publish a price — we have one builder for both, with prices on the page. VWO warns that changing traffic mid-test will “massively skew” results — we let you change traffic live with sticky assignment keeping every visitor in their variant. ABsmartly is code-only at €60K/yr — our client-side path needs no engineers. And none of them shows both statistical engines on one screen. We do. The full breakdown — where Optimizely still wins and when it is the right choice — is on the ProveBetter vs Optimizely comparison page.
Do I need developers?
For client-side tests, no: author variants in the builder, paste the generated snippet once, then edit the copy in the dashboard from then on. Server-side tests take a few lines of API code — the SDK is deliberately small (about 1.4 KB boot script, about 4.2 KB full SDK gzipped) and documented. Either way, a first test is designed to take under five minutes, not a procurement cycle.
Will my results stay valid if I change traffic mid-test?
Validity is engineered in, not assumed. Sticky hashing keeps every visitor in their variant, so nobody flips buckets partway when you move the dial. And the SRM check runs against the configured split, so if assignment ever drifts off-ratio the results page flags it before you can ship a bad decision. What we won't do is pretend a change never happened — silent re-bucketing is exactly the trap that made other platforms forbid mid-test changes.
What stats do you use?
Both families, and we print the method on every result. Frequentist: two-proportion z-test giving p-value, confidence level, and confidence interval, with multiple-comparison correction across variants. Bayesian: 95% credible interval, chance-to-beat-control, and expected loss — “what am I risking if I decide right now?”. Every verdict has a plain-language line. We're honest about peeking, too: early reads inflate false positives 5–10× at 10,000 observations (Johansson et al., KDD 2017), so the dashboard tells you when an early number isn't a decision yet. For the plain-English version — one test read both ways — Bayesian vs frequentist explainer.
What happens to my data?
It's yours. Every exposure and conversion is stored as a row in Postgres, readable through the API, and downloadable in one click — results and the raw event stream as CSV or JSON, free on every tier, with a documented, versioned schema (/export-schema.md). No export gate, no hostage data. The SDK uses a first-party cookie plus an opt-in consent gate: no silent fingerprinting, no third-party-cookie dependency.

Stop guessing. Start proving.

Create a test in minutes: pick your variants, split the traffic, and let two statistical engines agree on the answer.