Use when the user wants a VALID experiment instead of a guess — hypothesis, one variable, sample size and runtime math, statistical significance, primary versus secondary metrics, multi-arm designs, and a results template, across Optimizely, VWO, and native Meta and Google tests. Trigger on 'A/B test', 'split test', 'how long should I run the test', 'is this result significant', 'test two versions', 'which creative is actually better'. Also use when a winner was declared after two days on tiny numbers. Not for — scaling the proven winner, see `55-scaling-ads-global`; analyzing data already collected, see `13-data-analysis-global`; auditing the account, see `21-ads-audit-global`.
Installer avec Codex ou Claude Copiez ce prompt, collez-le dans Codex, Claude ou un autre assistant, puis laissez-le vérifier la page du skill et l'installer pour vous.
Une commande directe contourne le prompt de vérification. Examinez la source avant de l'exécuter.
Use when the user wants a VALID experiment instead of a guess — hypothesis, one variable, sample size and runtime math, statistical significance, primary versus secondary metrics, multi-arm designs, and a results template, across Optimizely, VWO, and native Meta and Google tests. Trigger on 'A/B test', 'split test', 'how long should I run the test', 'is this result significant', 'test two versions', 'which creative is actually better'. Also use when a winner was declared after two days on tiny numbers. Not for — scaling the proven winner, see `55-scaling-ads-global`; analyzing data already collected, see `13-data-analysis-global`; auditing the account, see `21-ads-audit-global`.
Run experiments that produce decisions, not noise. Most "A/B tests" in marketing are underpowered, peeked-at, and badly hypothesized — meaning the team learns nothing and ships the louder variant.
For Newbies
A valid A/B test answers one question: "Did this change cause a real improvement, or am I seeing noise?"
To answer it credibly you need four things:
A specific hypothesis with a numeric prediction
One variable changed (everything else identical)
Enough sample to detect the effect you care about
Statistical significance before you call a winner (typically p < 0.05)
If any one of these is missing, you don't have an A/B test — you have a coin flip with extra steps.
Common newbie mistake: running a test for 3 days, seeing variant B 40% higher, declaring victory, and shipping. Three days is too short to absorb day-of-week effects, and small samples produce wild swings. Variant B may revert (or reverse) by day 14.
Step 0 — Read Context
Read .agents/product-marketing-context.md if it exists. Audience size, average traffic, and current conversion rate determine whether a test is even feasible.
Step 1 — Information Gathering
Ask up to 4 questions:
What are you testing? (Ad headline / Landing page section / Email subject / Pricing display / CTA button / Creative video)
Daily traffic to the test surface? (Needed for sample size and duration)
Goal of the test? (Lift X% on primary metric / Pick a winner among N candidates / Validate a strategic hypothesis)
The 7 Principles of a Valid A/B Test
1. Test exactly one variable
The cardinal rule. Change two things at once and you cannot attribute the result.
Bad: "I changed the headline, the hero image, and the CTA color." → You learn nothing about which element drove the lift.
Good: Change only the headline. Image, CTA, layout, traffic source, and audience targeting are identical.
If you must test multiple changes, use a multivariate test (MVT) — but those need much more traffic (often 4×–8× a single A/B).
2. Hypothesize with a number
"If we [change X], [metric Y] will increase by [Z%] because [reason]."
Format:
Good: "If we change the CTA from 'Sign up' to 'Get my free demo,' conversion rate will increase by 15% because action-specific language reduces ambiguity."
Bad: "The new copy will be better." (No metric, no number, no causal reasoning — un-testable.)
The "because" matters: if your hypothesis is wrong but the reasoning was sound, you've still learned something generalizable.
3. Sufficient sample size
Don't stop early. Statistical tests need adequate data to distinguish signal from noise.
Minimum rule of thumb: 100 conversions per variant (not 100 visitors)
Better: Calculate sample size up front based on baseline conversion rate and minimum detectable effect (formula below)
4. Sufficient duration
Run for whole weeks, not 3 days, not 10 days. Different weekdays produce different audience behavior — Monday B2B traffic is not Saturday DTC traffic.
Minimum: 7 days
Recommended: 14 days
Watch for: holidays, paydays, monthly billing cycles, ad spend ramp-ups
5. Don't peek
Looking at results every hour and stopping when "B looks good" is the most common error in marketing experimentation. Early peeks combined with early stops dramatically inflate false positive rates.
Define the end date in advance. Honor it.
If you must monitor, use sequential testing methods designed for it (Bayesian frameworks like Optimizely's Stats Engine, or platforms with built-in sequential controls).
6. Statistical significance: p < 0.05
Most marketing teams use 95% confidence (p-value < 0.05) as the bar.
p-value < 0.05 → less than 5% chance the observed difference is random
p-value 0.05–0.10 → suggestive but inconclusive — extend the test
p-value > 0.10 → no evidence of an effect — keep control or test something else
Total: 9,600 sends per email — usually achievable in one campaign.
Feasibility quick-reference
Daily volume
Conv. rate
Days needed
Test feasibility
< 100
any
2+ months
Skip — focus on traffic first
100–500
2–5%
3–6 weeks
Yes, but be patient
500–2K
2–5%
2–3 weeks
Yes — ideal range
2K–10K
2–5%
1–2 weeks
Yes — rapid iteration
10K+
any
days
Yes — multi-arm tests possible
If volume is below 100/day, A/B testing is statistically wasted — concentrate on increasing traffic before running experiments.
Multi-Arm and Multivariate
Beyond simple A vs B:
Multi-arm (A/B/C/D): test 3+ variants at once. Sample size grows roughly linearly with arms.
Multivariate (MVT): test multiple elements simultaneously (headline × image × CTA = 8 combinations). Sample size grows multiplicatively. Only viable with very high traffic.
Sequential / Bayesian (Thompson Sampling): dynamically allocate more traffic to better-performing variants. Optimizely, Google Optimize successors, and Meta's auto-optimization use this.
For most teams: stick to A/B until traffic exceeds ~10K/day on the test surface.
What to Test (in priority order)
1. Headline — highest impact
Roughly 80% of visitors read the headline; 20% read the body. Optimizing the headline gives the largest expected lift per unit of effort.
Variations to try:
Question vs statement
Specific number vs generic ("3,247 founders trust us" vs "Trusted by founders")
Outcome-focused vs feature-focused
Short (5–7 words) vs long (12–15 words)
2. CTA button
Easy to change, often 5–25% lift potential.
Variations:
Text: "Sign up" vs "Get free demo" vs "Start free trial" vs "See pricing"
Color: brand primary vs contrast (high-contrast usually wins)
Size: standard vs large
Position: above-the-fold vs sticky vs end-of-page
3. Hero visual
Product shot vs lifestyle shot
Static image vs video
Founder face vs anonymous model
Demo screencast vs testimonial clip
4. Pricing display
Monthly vs annual primary
Strikethrough discount vs clean price
Number formatting ($299 vs $299.00 vs $299/mo)
Anchor pricing (3-tier with middle highlighted)
5. Social proof placement
Numbers vs detailed reviews
Logo wall vs customer count
Above CTA vs below CTA
Video testimonial vs text testimonial
6. Form fields
3 vs 5 vs 7 fields (fewer fields almost always wins on conversion, but lead quality may drop)
Label position (above vs left)
Single-step vs multi-step
Optional fields marked vs required marked
7. Email subject lines
Question vs benefit
Personalization vs generic
Emoji vs no emoji (varies by region/audience)
Length (under 40 chars vs 60+)
8. Ad creative
First 3-second hook variants
UGC style vs polished brand
Format: single image vs carousel vs video vs Reels-native
Headline on creative vs in copy field
Tool Recommendations
Tool
Best for
Cost
Meta Ads built-in A/B test
Creative, audience, placement on Meta
Free
TikTok Ads Split Test
TikTok ad creative and audience tests
Free
Google Ads Experiments
Google Ads campaigns and ad copy
Free
Optimizely Web
Enterprise web experimentation, sequential testing
$$$ enterprise
VWO
Mid-market web A/B + heatmaps
$199+/mo
Convert.com
Privacy-first web testing
$99+/mo
PostHog
Product feature flags + experiments + analytics
Free tier, generous
GrowthBook
Open-source A/B testing platform
Free / hosted plans
Statsig
Product experimentation with feature flags
Free tier
AB Tasty
Web experimentation + personalization
$$$
Unbounce / Instapage
Built-in A/B for landing pages
$90+/mo
Custom (split URL)
Two pages, 50/50 redirect, GA4/Pixel attribution
Free
Note: Google Optimize was sunset in September 2023. Migration paths: GA4 + a third-party platform (Optimizely, VWO, Convert) or PostHog/GrowthBook for product-led teams.
Setup Without Dedicated Tools
1. Build two versions of the page: /landing-a and /landing-b
2. Split traffic 50/50:
- Meta Ads: 2 ad sets, identical audience, different destination URLs
- Google Ads: 2 ads in the same ad group, identical targeting, different URLs
- Email: list-split feature in your ESP
3. Track conversions per variant:
- Meta Pixel custom event with parameter: page_version = "A" / "B"
- GA4 event with custom dimension
- PostHog feature flag exposure event
4. Run for the planned duration. Don't peek mid-test.
5. Export raw counts. Run significance test (calculator below).
Survey Monkey calculator — for sample size pre-test
Inputs:
Variant A: visitors + conversions
Variant B: visitors + conversions
Outputs:
p-value (need < 0.05 for 95% confidence)
Confidence interval on the lift
Lift % (relative or absolute)
Decision matrix
p-value
Lift size
Decision
< 0.05
> 5%
B wins — implement and document
< 0.05
< 5%
Significant but small — weigh implementation cost
0.05–0.10
> 10%
Borderline — extend test if feasible
> 0.10
any
No evidence — keep A or design a stronger test
Common Pitfalls
Peeking and stopping early. Most common cause of false positives. If the platform shows "B is winning" on day 3, the platform is misleading you (unless it's specifically designed for sequential testing).
Uneven splits. If split is 30/70 instead of intended 50/50, your delivery infrastructure has a bug. Investigate before trusting results.
Seasonality. Tests run only on weekdays vs weekends produce different results. Always run for whole weeks.
Novelty effect. New variants attract attention for the first 2–3 days, then performance regresses. Long enough tests absorb this.
Sample ratio mismatch (SRM). Even with 50/50 intent, if visitor counts diverge significantly (e.g. 8,400 vs 11,600), there's likely a tracking or assignment bug. Tools like PostHog and Optimizely flag this automatically.
Mixing traffic sources mid-test. Don't add a new ad campaign halfway through — it changes audience composition.
Multiple comparison problem. Running 20 simultaneous tests means ~1 will look "significant" by chance alone. Adjust thresholds (Bonferroni correction) or pre-register hypotheses.
Output Template
# A/B Test: [test name]
Created: [YYYY-MM-DD]
Owner: [name]
## 1. Hypothesis
"If we [change X], [metric Y] will increase by [Z%] because [reason]."
## 2. Variants- Variant A (Control): [current state description]
- Variant B (Challenger): [changed state description]
- Single change: [the one element that differs]
## 3. Metrics- Primary: [e.g. conversion rate]
- Secondary (guardrails): [e.g. bounce rate, time on page, AOV]
## 4. Sample size & duration- Baseline (p): [%]
- Minimum detectable effect (MDE): [%]
- Sample needed per variant: [N]
- Daily traffic to test surface: [N]
- Estimated days to complete: [N]
## 5. Setup- Tool: [Optimizely / VWO / PostHog / Meta built-in / custom]
- Variant A URL or asset: [...]
- Variant B URL or asset: [...]
- Tracking events: [list]
- Split ratio: 50/50
## 6. Timeline- Start: [date]
- End (planned): [date]
- Review meeting: [date]
## 7. Results (filled in after test ends)
| Variant | Visitors | Conversions | Rate | Lift vs A |
|---------|----------|-------------|------|-----------|
| A | | | | — |
| B | | | | +X% |
p-value: [x]
95% CI on lift: [lower%, upper%]
Significant (p < 0.05): [Yes / No]
## 8. Decision
[Ship B / Keep A / Inconclusive — extend or redesign]
## 9. Action
[Implement variant B globally / Roll back / Schedule next iteration]
## 10. Lessons
[What this teaches generalizable for future tests]
Quality Checklist
Exactly one variable changed
Hypothesis is specific and includes a numeric prediction
Sample size calculated up front (minimum 100 conv/variant)
Duration is at least 1 full week, ideally 2 weeks
No peeking — end date defined and honored
p-value calculated before declaring a winner (target p < 0.05)
Result documented (winner or not — both are learning)
Secondary metrics checked (no guardrail violations)
Sample ratio verified (no SRM red flags)
Next test identified based on what this one taught
