ab-test-setup

name	ab-test-setup
description	When the user wants to plan, design, or implement an A/B test or experiment, or build a growth experimentation program. Also use when the user mentions 'A/B test,' 'split test,' 'experiment,' 'conversion test,' or 'growth experiment.'
allowed-tools	Read, Write, WebSearch, WebFetch, AskUserQuestion
model	sonnet

A/B Testing & Experimentation

Expert knowledge for designing statistically valid experiments that produce trustworthy, actionable results.

Core Framework: The Experimentation Stack

1. Hypothesis → 2. Metric Selection → 3. Sample Size → 4. Test Design → 5. Run → 6. Interpret → 7. Act

Every step must be completed before the next. Skipping hypothesis formation produces uninterpretable results.

1. Hypothesis Formation

Template: "Because [evidence/observation], we believe [change] will [metric direction] for [audience segment], measured by [primary metric]."

Example: "Because our heatmaps show 60% of users ignore our hero CTA, we believe moving it above the fold will increase free trial signups for new visitors, measured by signup rate."

Good hypotheses have:

• A root cause (not just "let's try X")
• A directional prediction (increase/decrease, not just "improve")
• A named primary metric
• An audience segment

Bad hypothesis signals: "Let's test a new headline," "Maybe color X converts better," "Users might prefer Y."

2. Metric Selection

Primary metric: One metric that defines success. The test either wins or loses on this.

Secondary metrics: Supporting signals that explain why the primary moved.

Guardrail metrics: Metrics that must not regress. If they do, the test fails regardless of primary metric lift.

Metric Type	Example	Purpose
Primary	Trial signup rate	Win/loss decision
Secondary	CTA click rate, page scroll depth	Diagnose cause
Guardrail	Bounce rate, page load time	Prevent regressions

Avoid: Multiple primary metrics. If both need to improve, run two tests.

3. Sample Size Calculation

Required inputs:

• Baseline conversion rate (current)
• Minimum Detectable Effect (MDE) — smallest improvement that justifies the change
• Statistical significance threshold (α) — typically 0.05
• Statistical power (1-β) — typically 0.80

Rule of thumb formula:

n = (16 × σ²) / δ²

Where σ = standard deviation, δ = minimum detectable difference.

Online calculators: Evan Miller's A/B test calculator, Optimizely's calculator.

Typical sample sizes for conversion rate tests:

• Baseline 2%, MDE 20% relative → ~10,000 visitors per variant
• Baseline 5%, MDE 10% relative → ~20,000 visitors per variant
• Baseline 10%, MDE 5% relative → ~50,000 visitors per variant

Critical: Calculate sample size BEFORE running. Post-hoc power analysis is invalid.

4. Test Design Decisions

Frequentist vs Bayesian

Frequentist (Classical):

• Fixed sample size, binary decision at end
• Uses p-values and confidence intervals
• Best for: regulated industries, teams that need clear "yes/no"
• Risk: peeking at results invalidates the test

Bayesian:

• Continuous probability updates
• Uses probability of being best / expected loss
• Best for: fast-moving teams, partial rollouts
• Allows stopping early without inflating false positive rate

Recommendation: Use Bayesian when traffic is limited or you need to stop early. Use frequentist when stakes are high and you can commit to the full run.

A/B vs Multivariate vs Multi-Armed Bandit

A/B (or A/B/n): Test one variable across 2+ variants.

• Use when: you have a clear hypothesis about one element
• Traffic split: even across all variants

Multivariate (MVT): Test multiple variables simultaneously in combinations.

• Use when: you want to understand interaction effects
• Requires: significantly more traffic (n × m variants)
• Risk: very few teams have enough traffic for valid MVT

Multi-Armed Bandit: Adaptive allocation — more traffic to better-performing variants over time.

• Use when: you cannot afford to run a losing variant for long
• Trade-off: explores less, so may miss global optimum

Segmentation & Personalization Tests

Segment-level analysis is fine AFTER a test completes. But do not start a test with the goal of "finding the best variant for segment X" unless you've pre-specified that segment and sized accordingly (Bonferroni correction applies).

5. Test Duration

Never stop early based on results. Run for:

• At least the calculated sample size
• At minimum 2 full business cycles (usually 2 weeks) to account for day-of-week effects
• Until weekly variation stabilizes

Business cycle effects to account for:

• Day-of-week (B2B traffic often drops on weekends)
• Payday effects (consumer purchases spike)
• Seasonal events
• Paid campaign start/end dates

6. Result Interpretation

Statistical Significance ≠ Practical Significance

A result can be statistically significant (p < 0.05) but practically meaningless (0.1% lift). Always check:

1. Is the p-value below your threshold?
2. Is the confidence interval entirely above your MDE?
3. Did guardrail metrics hold?

The Novelty Effect

New variants often see an initial uplift from curiosity, then regress. If your test duration is short, results may not hold post-launch. Prefer tests that run for 4+ weeks.

Peeking Problem

Checking results during a test and stopping when you see p < 0.05 inflates false positive rate dramatically. With α = 0.05, checking at every 10% of collected data gives an actual false positive rate of ~40%.

Solutions:

• Pre-commit to sample size before starting
• Use sequential testing / always-valid p-values (Johari et al.)
• Use Bayesian methods with proper stopping rules

Common Result Scenarios

Scenario	What It Means	Action
Winner, practical significance	Test worked	Ship variant, document learnings
Winner, no practical significance	Stat artifact or tiny effect	Do not ship unless free; learn from it
No winner	No effect detected	Bigger change needed, or hypothesis wrong
Variant loses	Baseline is better	Keep control, investigate why
Guardrail violated	Variant causes harm	Stop test, keep control

7. Building an Experimentation Program

Test Prioritization: PIE Framework

Score each test idea on 3 dimensions (1-10):

• Potential: How much improvement is possible?
• Importance: How important is the page/flow?
• Ease: How easy is it to implement?

PIE score = (P + I + E) / 3. Run highest scores first.

Velocity Targets

• Prototype stage: 1-2 tests/month is fine
• Pilot stage: 2-4 tests/month
• Candidate/Production: 5+ concurrent tests if traffic allows

Test Taxonomy (track in a log)

For every test, record:

• Hypothesis
• Variants
• Primary metric + result
• Secondary metrics
• Duration + sample size
• Decision made
• Learning extracted

Winning Variants Compounding

Each winning test locks in. Subsequent tests build on previous winners. This is the compound growth engine of CRO.

Common Rationalizations

Rationalization	Reality
"We don't have enough traffic — let's just run it for a week"	Underpowered tests produce noise, not signal. Calculate the required duration first.
"The results look good at day 3, let's stop early"	Early stopping with p < 0.05 produces a ~40% false positive rate. Run to completion.
"We're testing 5 things at once on this page"	Multivariate tests need 5–10× more traffic. You almost certainly don't have it.
"Statistical significance is all we need"	Stat sig tells you the effect is real. Practical significance tells you if it matters. Check both.
"Let's test everything and see what works"	No hypothesis = no learning. Even a negative result teaches nothing without a hypothesis.
"Our audience is different — 80% confidence is fine"	Lowering your threshold increases false positive rate proportionally. There are no free lunches.

Verification

• Hypothesis written using the template (evidence → change → metric direction → audience → measurement)
• Primary metric defined (singular)
• Guardrail metrics defined
• Sample size calculated before test launch with baseline rate and MDE specified
• Test duration set to cover minimum 2 business cycles
• No mid-test peeking protocol agreed upon
• Result interpretation includes practical significance check (not just p-value)
• Test logged with hypothesis, result, and learning extracted