ab-test-generator v1.0.0A

Immutable. This exact content is served forever at /api/v1/blob/9a6ca7df76f5aa53.

---
name: ab-test-generator
description: "Reads page analytics and heatmap data from Humblytics, generates A/B test hypotheses with element selectors, and launches no-code split tests via the Humblytics API. Use when creating A/B tests, split tests, multivariate tests, or when you need to test headlines, CTAs, layouts, or pricing. Triggers: A/B test, split test, experiment, test hypothesis, launch test, variant."
metadata:
  version: 1.0.0
  author: Humblytics
---

# A/B Test Generator

## Purpose

Generate data-driven A/B test configurations from Humblytics analytics and heatmap data. This skill creates complete test definitions including hypotheses, variant specifications with CSS/element selectors, success metrics, sample size requirements, and can launch tests directly through the Humblytics API.

## When to Use

- Creating A/B tests from conversion data or heatmap insights
- Generating test hypotheses for a specific page or funnel step
- Launching no-code split tests via the Humblytics testing API
- Calculating required sample size and test duration
- Designing multivariate test matrices
- Reviewing and iterating on existing test results

## Credentials

This skill reads a Humblytics API key from the environment. **Never paste API keys directly into chat** — they persist in transcripts and logs.

Setup (one time):
1. `cp .env.example .env` at the repo root and fill in `HUMBLYTICS_API_KEY`
2. `source .env` in your shell before running the agent (or use `direnv`, or add the exports to your shell profile)
3. Get the key from Humblytics Dashboard > Settings > API
4. The skill will ask for your **Property ID** (also in Dashboard > Settings > API)

- **Base URL**: `https://app.humblytics.com/api/external/v1`
- **Docs**: https://docs.humblytics.com/api

If `HUMBLYTICS_API_KEY` is not in the environment, stop and point the user at `.env.example` — do not accept the key in chat.

## Before You Start

1. **Confirm property and page** — Which Humblytics property and which page URL to test
2. **Pull current data** — Retrieve page analytics, heatmap data, and current conversion rate
3. **Check existing tests** — Look for any running tests to avoid conflicts
4. **Understand the goal** — What is the primary conversion action on this page?
5. **Verify traffic volume** — Ensure enough traffic for statistical significance within a reasonable timeframe
6. **Check for context** — Look for product briefs, AGENTS.md, or existing CRO documents that inform test direction

## Core Workflow

### Step 1: Gather Page Intelligence

Pull data from Humblytics:

- **Page analytics**: Traffic volume, bounce rate, time on page, scroll depth
- **Heatmap data**: Click maps, scroll maps, attention maps
- **Event data**: CTA clicks, form interactions, video plays
- **Device split**: Mobile vs desktop behavior differences
- **Source split**: How different traffic sources behave on this page

API endpoints (all under `/api/external/v1/`, all take `start`, `end`, `timezone`):
- `GET /properties/{propertyId}/pages/details?page=/path` — Single-page deep dive (UTM, device, country, scroll depth, bounce)
- `GET /properties/{propertyId}/clicks/details?page=/path` — Click data with UTM attribution (the public API has no `/heatmaps` endpoint; click data is the closest analogue)
- `GET /properties/{propertyId}/forms/details?page=/path` — Form submissions and conversion rates for that page (no generic events endpoint exists publicly)
- `GET /properties/{propertyId}/split-tests` — List existing experiments. Optional `?status=active|complete`

### Step 2: Identify Test Opportunities

Analyze the data for signals:

**High-value signals from heatmaps:**
- Users clicking non-clickable elements (rage clicks) — make them clickable or remove confusion
- Low scroll depth — critical content is below the fold, move it up
- CTA getting few clicks despite visibility — copy, color, or placement issue
- Form field abandonment — simplify or reorder fields
- Dead zones — large page areas with no interaction

**High-value signals from analytics:**
- High bounce rate from specific sources — message mismatch
- Mobile conversion significantly lower than desktop — responsive layout issue
- High time-on-page but low conversion — users are interested but not persuaded
- Low time-on-page and low conversion — page fails to engage

### Step 3: Formulate Hypotheses

For each opportunity, create a structured hypothesis:

```
Test Name: [descriptive-slug]
Page: [URL]
Hypothesis: IF we [specific change], THEN [primary metric] will [direction]
            BECAUSE [evidence from data]

Control: [current state description]
Variant: [proposed change description]

Primary Metric: [conversion event or goal]
Secondary Metrics: [engagement metrics to monitor]

Element Selector: [CSS selector for the element to modify]
Change Type: [text | style | visibility | layout | redirect]
```

### Step 4: Calculate Sample Size and Duration

For each test, calculate statistical requirements:

**Inputs needed:**
- Baseline conversion rate (from current analytics)
- Minimum detectable effect (MDE) — typically 10-20% relative lift
- Statistical significance level — default 95%
- Statistical power — default 80%

**Sample size formula (per variant):**
```
n = (Z_alpha/2 + Z_beta)^2 * (p1(1-p1) + p2(1-p2)) / (p2 - p1)^2
```

**Duration estimate:**
```
Days = (n * number_of_variants) / daily_traffic_to_page
```

Present this clearly:
- Required sample per variant
- Total sample needed
- Estimated days to reach significance at current traffic
- Recommendation: proceed, increase traffic first, or test a larger change

### Step 5: Define Test Configuration

Create the test spec to send to `POST /properties/{propertyId}/split-tests`. The required body shape is:

```json
{
  "name": "descriptive-test-name",
  "page": "/pricing",
  "type": "nocode",
  "variants": [
    { "label": "control", "changes": [] },
    {
      "label": "variant-a",
      "changes": [
        {
          "selector": "#hero-cta",
          "attribute": "textContent",
          "value": "See Your Analytics"
        }
      ]
    }
  ],
  "goal": "signup_started",
  "auto_stop_days": 30
}
```

Required fields: `name`, `page`, `type` (use `"nocode"` for selector-based tests), `variants` (each with `label` + `changes`).
Optional: `goal` (primary conversion event), `auto_stop_days` (auto-end the test after N days).

### Step 6: Launch or Document

**To launch via API:**
- `POST /properties/{propertyId}/split-tests` — Create and start the test (body shape above)
- `GET /properties/{propertyId}/split-tests/{experimentId}` — Experiment details with per-variant metrics inline (no separate `/results` endpoint — variant metrics come back in the same response)
- `PATCH /properties/{propertyId}/split-tests/{experimentId}` — Update an active experiment. Body: `{ "name": "...", "auto_stop_days": N }`
- `POST /properties/{propertyId}/split-tests/{experimentId}/stop` — Stop a running experiment. Body: `{ "reason": "..." }`
- `GET /properties/{propertyId}/split-test-recommendations?page=/path` — AI-generated split-test suggestions for a page

**To document for manual launch:**
- Output the full test specification
- Include screenshot annotations if click-data informed the test
- Provide the hypothesis document for the team

## Test Type Selection Guide

| Scenario | Test Type | Notes |
|----------|-----------|-------|
| One element change | A/B test | Fastest to significance |
| Two element changes | A/B/C test | Test both independently |
| Multiple interacting elements | Multivariate | Needs 4x+ traffic |
| Completely different page | Split URL test | Redirect-based |
| Copy variations only | A/B test | Quick wins |
| Layout restructure | Split URL test | Build separate variant page |

## Common Test Categories

### Headlines and Copy
- Value prop rewording
- Specificity (add numbers, timeframes, outcomes)
- Emotional vs rational framing
- Length (short punchy vs detailed)

### CTAs
- Button text (action verbs, benefit language, urgency)
- Button color and size
- Button placement (above fold, after social proof, sticky)
- Single CTA vs multiple CTAs

### Social Proof
- Testimonials vs logos vs metrics
- Placement (near CTA vs header vs throughout)
- Specificity (named customers vs anonymous)

### Layout and Structure
- Long page vs short page
- Information hierarchy reordering
- Form length and field order
- Navigation presence vs removal

### Pricing
- Price anchoring (show higher price first)
- Plan naming
- Feature comparison layout
- Free trial vs freemium vs demo

## Avoiding Common Mistakes

1. **Testing too small a change** — If your MDE requires 50,000 visitors per variant, the change is too subtle. Test bolder.
2. **Running too many tests on one page** — Interaction effects corrupt results. One test per page at a time.
3. **Stopping early on positive results** — Peeking inflates false positives. Commit to the sample size.
4. **Ignoring secondary metrics** — A variant that increases signups but increases churn is not a win.
5. **Not segmenting results** — A test can be flat overall but show strong wins on mobile. Always segment.

## Output Format

For each generated test, present:

1. **Hypothesis** — One sentence: IF/THEN/BECAUSE
2. **Evidence** — What data supports this test
3. **Variants** — Control and variant descriptions
4. **Selectors** — CSS selectors and exact changes
5. **Metrics** — Primary and secondary goals
6. **Duration** — Estimated days to significance
7. **Expected Impact** — Projected conversion lift range

## Related Skills

- **cro-optimizer** — Identify which pages and funnel steps need testing
- **page-cro** — Deep page-level audit to inform test hypotheses
- **copywriting** — Generate high-quality copy variants for tests
- **funnel-reporter** — Track how test results affect downstream funnel metrics