A/B Test Hypothesis Generator: 10 Tests Worth Running
A backlog of "change the button color" experiments is how growth teams stall. This AI tool starts from the real drop-off in your funnel and generates 10 hypotheses across the levers that matter — copy, hierarchy, social proof, pricing, form design, CTA — each structured for clean measurement.
Full Prompt
Generate a ranked list of 10 A/B test hypotheses for a specific page or flow, each structured so a growth team can prioritize, run, and learn from it cleanly. HYPOTHESIS METHODOLOGY (follow in order): 1. Diagnose the Drop-Off Goal: Test where the friction actually is. - Restate the page or flow and the conversion goal. - From the data provided, identify the steepest drop-off or biggest friction point. - Note any qualitative signal (heatmaps, session recordings, support tickets). 2. Generate Across Levers Cover at least 5 of these levers: - Headline / hero copy - Hierarchy and information architecture - Social proof (placement, type, volume) - Pricing presentation and anchoring - Form design (fields, steps, defaults) - Visual proof (screenshots, video, demo) - CTA copy and placement - Loading and perceived performance 3. Structure Each Hypothesis Use the format: - We believe [change] will cause [metric] to [direction] because [reason grounded in user behavior]. - We'll know it worked if [primary metric] moves by [magnitude] over [duration / sample size]. 4. Rank by Impact and Effort For each: - Expected impact (Low / Med / High) - Effort to build (Low / Med / High) - Confidence based on existing evidence (Low / Med / High) - Pick the top 3 by ICE score. OUTPUT CONSTRAINTS: - Return exactly 10 hypotheses. - Every hypothesis ties to a specific behavior or data point — no random "change the button to red." - Highlight the top 3 in a separate block with the recommended test order. - Flag any test that needs more traffic than the page can realistically deliver. --- MY INFO: Page or Flow (required): [URL or description] Primary Conversion Metric (required): Current Conversion Rate and Volume (required): What You've Already Tested (optional): Qualitative Signal (optional): [heatmaps, session recordings, complaints]
What You Get
- 10 hypotheses across at least 5 different levers
- A standard format — "We believe X will cause Y because Z; we'll know it worked if..."
- An ICE score (Impact, Confidence, Effort) for each
- The top 3 in recommended run order with rationale
Why It Works
Every hypothesis ties to a specific behavior or data point — random ideas get rejected. The format forces a primary metric, expected magnitude, and a duration or sample size, so the test produces a real answer rather than another inconclusive result. Tests that would need more traffic than the page realistically delivers get flagged before they consume a quarter.
Best Practices
- Show the data: A current conversion rate and traffic volume changes which tests are even viable.
- Bring qualitative signal: A heatmap or session recording beats imagination.
- Don't test five things at once: One change per arm; one primary metric per test.
- Power-check it: Underpowered tests aren't tests — they're guesses with extra steps.
Run the tests that move the metric and skip the ones that move the meeting.