5 Prompts to Generate A/B Test Hypotheses with AI

← All articles

The best A/B tests start with a well-formed hypothesis: a specific, testable prediction about what will change, and why. Generating a strong hypothesis backlog used to take days of analysis. With AI, it takes an afternoon — if you know how to prompt.

One illuminated lightbulb standing out among unlit ones — the winning test hypothesis

Most hypotheses won't win. The goal is to generate enough good ones that the one that does gets tested first.

Personalising the call to action delivers a 217.72% uplift on average — up from 181.35% in 2025. That's the highest single-tactic gain in the 2026 survey. Most teams aren't running it because they assume personalisation requires complex tooling. It doesn't. A strong hypothesis is the real starting point. — 2026 Growth Marketer Survey

The Hypothesis Template

Before the prompts, the structure: "We believe that [change] will [direction] [metric] for [audience] because [rationale]. We'll know this is true when [measurement criterion]."

Every prompt you write should produce output in this form or close to it. Vague outputs ("test a different headline") are not actionable. Specific outputs ("test a benefit-led headline naming the specific outcome for paid search visitors, because they're already problem-aware and need solution confirmation") are ready to execute.

Prompt 1 — Page Analysis

Use this when you want to audit a specific page for friction and test opportunities.

Here is the current copy for our landing page: [PASTE COPY] Our baseline conversion rate is [X]%. Our primary traffic source is [source]. Identify the top 5 friction points that may be preventing visitors from converting, and for each, write a specific A/B test hypothesis in this format: "We believe that [change] will [direction] [metric] because [rationale]."

Prompt 2 — Competitor Contrast

Use this to generate variants informed by what the rest of your market is doing.

Here are the headlines and CTAs from 5 competitor landing pages in our category: [PASTE LIST] Our current headline is: [PASTE HEADLINE] Identify what they're doing differently from us, and generate 5 variant headlines we could A/B test against the current one. For each, explain why it might outperform.

Prompt 3 — Persona Mismatch

Use this when your copy and your audience may be misaligned — common on pages that haven't been updated since product-market fit shifted.

Our landing page uses this copy: [PASTE] Our ideal customer is: [DESCRIPTION OF ICP — role, company size, primary pain point] Identify any places where the copy assumes knowledge, uses jargon, or doesn't address the ICP's primary concern. Generate 3 test hypotheses that close each gap.

Prompt 4 — Follow-On Hypotheses

Use this after a test concludes — win or loss — to generate the next round of ideas informed by what you just learned.

Our A/B test showed: [DESCRIBE RESULT — e.g., "benefit-led headline lifted conversion 28% over product-led headline, p=0.02"] Generate 5 follow-on test hypotheses that extend this result. Either apply the same principle to a different page element, or propose a test that would help explain WHY the variant won.

Prompt 5 — Objection Mapping

Use this to build tests that directly address conversion barriers rather than optimising what's already working.

What are the most common objections a visitor might have to [PRODUCT/OFFER]? For each objection, write a landing page element test hypothesis that directly addresses it. Include: what element to change, what the variant says, and why this objection-handling approach might improve conversion.

Use these prompts on a rotating basis — one per week — and you'll generate more testable hypotheses than most teams can execute in a quarter. The bottleneck shifts from ideation to prioritisation, which is the right problem to have.

Turn your hypotheses into live experiments

Iteratist lets you go from hypothesis to running test in minutes, with built-in significance tracking.

Start free →