qbrix
Platform
Platform · Experiments

The version that is losing should not keep getting more traffic.

A classic A/B test splits traffic evenly and holds it there until someone calls the result. qbrix moves traffic toward whichever version is doing better, automatically.

Read the reference
checkout-ctaRunning · day 14
OverviewArmsPolicyGate
Poolcheckout
Variants4
StrategyThompson Sampling
Reward rate6.42%
Decisions · 24h48,210
Variants in this pool
VariantShareReward
social-proof76.4%6.42%
control12.1%4.81%
urgency-copy8.4%4.02%
discount-badge3.1%3.44%

Split your traffic four ways. Three quarters of it goes to versions you are about to throw away, and it keeps going there until someone calls the test.

A/B test · even split
day 1day 14day 28
social-proofcontrolurgency-copydiscount-badge

Nothing moves until someone calls the test.

qbrix · same four weeks
day 1day 14day 28
social-proofcontrolurgency-copydiscount-badge

Traffic moves to what is working, while the test runs.

Write the versions once. Point anything at them.

A pool is your set of versions — the four checkout buttons. An experiment points at one and adds an audience, a strategy and a goal. Run a second for desktop without duplicating a word of copy.

1.0Structure
Pool · checkout
4 versions
social-proofcontrolurgency-copydiscount-badge
checkout-ctaRunning · day 14
Gatemobile · NL, BE, DE
StrategyThompson Sampling
checkout-cta-desktopRunning · day 6
Gatedesktop · all markets
StrategyUCB1-Tuned

You decide what winning means.

A click. A completed purchase. The size of the basket. A rating out of five. You send the number that matters to your business, and that is what qbrix optimises for.

2.0Reward
Did they act?reward = 1 or 0clicks, sign-ups, completed checkouts — anything with a yes and a no
How much did it bring in?reward = 84.00basket size, revenue, seconds on the page — the number itself
How good was it?reward = 0.72a rating, a quality score, anything you can put on a nought-to-one scale
And it does not have to arrive in the same request
14:32:07You ask, and render the answer→ social-proof · sel_9f2a41
nine minutes of shoppingyou store nothing — the request id is the link
14:41:02They check out→ reward 1.0 · sel_9f2a41

Let qbrix pick the best learning model for your problem.

How boldly traffic moves, and how long a losing version keeps being tried, is a choice. Make it yourself, or take the recommendation and move on — your app is written the same way either way.

3.0Strategy
New experiment
Basics·Reward type·Strategy·Feature gate
AutoRecommended

Tries several ways of learning side by side and moves traffic toward whichever is working best on your data.

Learns from your data and adapts toward the best strategy
AdvancedChoose a specific strategy instead
Back to Auto
Strategy
LinUCBPolicy
LinUCBPolicycontextual
binaryboundedcontinuous
What the choice actually decides
  • How boldly it commits to a front-runner early on
  • How long a losing version keeps being tried anyway
  • Whether it assumes your traffic behaves the same all week
  • Whether it learns one answer, or one per kind of visitor
and whichever you pick, your app is:qbrix.select(…) · qbrix.feedback(…)
See the strategies in the docs

Keep the control, while qbrix optimizes.

Everything about who takes part stays yours to move, any time, without a deploy.

4.0In flight
Change any time, without a deploy
How much traffic takes partMove the rollout from 1% to everyone. It takes effect within seconds.
Who is kept out of itAdd or edit the rules. Whoever you exclude keeps seeing your current version.
Whether it runs at allPause it, look at it, resume it. Nothing it has learned is lost.
What you send about a visitorAdd a property to the request whenever you have one worth sending.
Fixed when it starts
The set of versionsA pool's versions are settled at creation, so every number an experiment reports is about the same four things.
The strategy, and its starting settingsSwapping how it learns halfway through would quietly invalidate everything it has learned so far.
Changing your mind about either one is a new experiment — pointed at the same pool, and reached by the same two calls.
More in PlatformSee how it fits together
01Decisioning

The best experience for each visitor, on every request

03Feature gates

Who takes part, at what pace, and always the same answer

04Personalization

A different answer for each kind of visitor