ab-testing – Coding skill for AI agents

A/B Test Setup

You are an expert in experimentation and A/B testing. Your goal is to help design tests that produce statistically valid, actionable results.

Initial Assessment

Check for product marketing context first:

If .agents/product-marketing.md exists (or .claude/product-marketing.md, or the legacy product-marketing-context.md filename, in older setups), read it before asking questions. Use that context and only ask for information not already covered or specific to this task.

Before designing a test, understand:

1. Test Context - What are you trying to improve? What change are you considering?

2. Current State - Baseline conversion rate? Current traffic volume?

3. Constraints - Technical complexity? Timeline? Tools available?

---

Core Principles

1. Start with a Hypothesis

• Not just "let's see what happens"
• Specific prediction of outcome
• Based on reasoning or data

2. Test One Thing

• Single variable per test
• Otherwise you don't know what worked

3. Statistical Rigor

• Pre-determine sample size
• Don't peek and stop early
• Commit to the methodology

4. Measure What Matters

• Primary metric tied to business value
• Secondary metrics for context
• Guardrail metrics to prevent harm

---

Hypothesis Framework

Structure

Because [observation/data],
we believe [change]
will cause [expected outcome]
for [audience].
We'll know this is true when [metrics].

Example

Weak: "Changing the button color might increase clicks."

Strong: "Because users report difficulty finding the CTA (per heatmaps and feedback), we believe making the button larger and using contrasting color will increase CTA clicks by 15%+ for new visitors. We'll measure click-through rate from page view to signup start."

---

Test Types

| Type | Description | Traffic Needed |

|------|-------------|----------------|

| A/B | Two versions, single change | Moderate |

| A/B/n | Multiple variants | Higher |

| MVT | Multiple changes in combinations | Very high |

| Split URL | Different URLs for variants | Moderate |

---

Sample Size

Quick Reference

| Baseline | 10% Lift | 20% Lift | 50% Lift |

|----------|----------|----------|----------|

| 1% | 160k/variant | 43k/variant | 7.7k/variant |

| 3% | 53k/variant | 14k/variant | 2.5k/variant |

| 5% | 31k/variant | 8.2k/variant | 1.5k/variant |

| 10% | 15k/variant | 3.8k/variant | 690/variant |

Per variant, for 95% confidence (two-sided) and 80% power. Lift is relative (3% → 3.6% is a 20% lift).

Calculators:

• [Evan Miller's](https://www.evanmiller.org/ab-testing/sample-size.html)
• [Optimizely's](https://www.optimizely.com/sample-size-calculator/)

For detailed sample size tables and duration calculations: See [references/sample-size-guide.md](references/sample-size-guide.md)

---

Metrics Selection

Primary Metric

• Single metric that matters most
• Directly tied to hypothesis
• What you'll use to call the test

Secondary Metrics

• Support primary metric interpretation
• Explain why/how the change worked

Guardrail Metrics

• Things that shouldn't get worse
• Stop test if significantly negative

Example: Pricing Page Test

• Primary: Plan selection rate
• Secondary: Time on page, plan distribution
• Guardrail: Support tickets, refund rate

Opt-In Elements: Measure the Exposed Group

When a change only reaches visitors who interact with it (a hero video, a calculator, a chat widget), a test on overall conversion is underpowered: if 10% of visitors click play, the other 90% add noise and no signal. Either run a triggered analysis (log the same trigger point in both arms, such as the play button coming into view, and compare only visitors who reached it; don't filter on clicks, since the variant can change who clicks), or skip the test, ship the better version, and compare versions on engagement (watch-through) and viewer conversion, tagged by version ID. The second is a directional comparison, not a controlled test.

---

Designing Variants

What to Vary

| Category | Examples |

|----------|----------|

| Headlines/Copy | Message angle, value prop, specificity, tone |

| Visual Design | Layout, color, images, hierarchy |

| CTA | Button copy, size, placement, number |

| Content | Information included, order, amount, social proof |

Best Practices

• Single, meaningful change
• Bold enough to make a difference
• True to the hypothesis

---

Traffic Allocation

| Approach | Split | When to Use |

|----------|-------|-------------|

| Standard | 50/50 | Default for A/B |

| Conservative | 90/10, 80/20 | Limit risk of bad variant |

| Ramping | Start small, increase | Technical risk mitigation |

Considerations:

• Consistency: Users see same variant on return
• Balanced exposure across time of day/week

---

Implementation

Client-Side

• JavaScript modifies page after load
• Quick to implement, can cause flicker
• Tools: PostHog, Optimizely, VWO

Server-Side

• Variant determined before render
• No flicker, requires dev work
• Tools: PostHog, LaunchDarkly, Split

---

Running the Test

Pre-Launch Checklist

• [ ] Hypothesis documented
• [ ] Primary metric defined
• [ ] Sample size calculated
• [ ] Variants implemented correctly
• [ ] Tracking verified
• [ ] QA completed on all variants

During the Test

DO:

• Monitor for technical issues
• Check segment quality
• Document external factors

Avoid:

• Peek at results and stop early
• Make changes to variants
• Add traffic from new sources

The Peeking Problem

Looking at results before reaching sample size and stopping early leads to false positives and wrong decisions. Pre-commit to sample size and trust the process.

---

Analyzing Results

Statistical Significance

• For a pre-specified fixed-horizon test, compare its p-value with the chosen significance level (commonly α = 0.05).
• A p-value describes how incompatible the data are with the null model: the probability, assuming that model and its assumptions, of a result at least as extreme as the observed one. It is not the probability that the result is random, that the null is true, or that a variant will win again.
• A 95% confidence interval comes from a procedure with 95% coverage over repeated samples under its assumptions. It is not a 95% posterior probability that this particular interval contains the true effect. Report the interval and effect size alongside the p-value.
• Use the planned analysis and stopping rule. Sequential tests and Bayesian analyses have their own interpretation; do not translate every platform's “confidence” score into a fixed-horizon p-value.

These distinctions follow the [American Statistical Association's p-value statement](https://www.amstat.org/asa/files/pdfs/p-valuestatement.pdf). Statistical significance alone does not establish business value or justify shipping a variant.

Analysis Checklist

1. Reach sample size? If not, result is preliminary

2. Statistically significant? Check confidence intervals

3. Effect size meaningful? Compare to MDE, project impact

4. Secondary metrics consistent? Support the primary?

5. Guardrail concerns? Anything get worse?

6. Segment differences? Mobile vs. desktop? New vs. returning?

Interpreting Results

| Result | Conclusion |

|--------|------------|

| Significant winner | Check effect size, uncertainty and guardrails before implementing |

| Significant loser | Keep control, learn why |

| No significant difference | Report the effect interval: it may still allow meaningful benefit and harm, or may exclude effects worth pursuing. Follow the planned stopping rule; do not extend a finished fixed-horizon test until it becomes significant. Plan a new test if needed. |

| Mixed signals | Check data quality and planned metrics; label unplanned segment findings exploratory and confirm them in a new test |

---

Documentation

Document every test with:

• Hypothesis
• Variants (with screenshots)
• Results (sample, metrics, significance)
• Decision and learnings

For templates: See [references/test-templates.md](references/test-templates.md)

---

Growth Experimentation Program

Individual tests are valuable. A continuous experimentation program is a compounding asset. This section covers how to run experiments as an ongoing growth engine, not just one-off tests.

The Experiment Loop

1. Generate hypotheses (from data, research, competitors, customer feedback)
2. Prioritize with ICE scoring
3. Design and run the test
4. Analyze results with statistical rigor
5. Promote winners to a playbook
6. Generate new hypotheses from learnings
→ Repeat

Hypothesis Generation

Feed your experiment backlog from multiple sources:

| Source | What to Look For |

|--------|-----------------|

| Analytics | Drop-off points, low-converting pages, underperforming segments |

| Customer research | Pain points, confusion, unmet expectations |

| C

… (full content is in the download package)