How to Run a Curiosity Gap Email Test
The curiosity gap is one of the oldest tricks in subject line writing: hint at something valuable without revealing it, so readers open to close the loop. The hard part isn't writing the line — it's proving it actually beats your control. This guide shows how to run a curiosity gap email test with marginal, a hosted email MCP server your AI agent can drive directly.
What a curiosity gap email test measures
A curiosity gap subject line withholds the payoff — "You're probably making this one mistake" instead of "3 ways to fix your onboarding flow." It tends to lift open rates, but it can also burn trust or hurt clicks if the body doesn't deliver. A proper test isolates the curiosity-gap variant against a clear, descriptive control so you see the real trade-off.
That means your curiosity gap email test should track both opens and clicks, not opens alone. marginal records open and click tracking per variant and computes within-test lift, so you can tell whether the intrigue translated into action or just inflated the open number.
- Control: a plain, descriptive subject line
- Variant: the curiosity-gap version that hides the payoff
- Metrics: open rate, click rate, and within-test lift
- Goal: confirm the gap drives clicks, not just opens
Running the test with the marginal email MCP server
marginal exposes four MCP tools to an AI agent: generate_variants, launch_test, get_results, and recommend_next. From a client like Cursor, Claude Code, or Windsurf, you can ask the agent to draft a curiosity-gap subject line set, ship the subject line A/B test, and report back — without leaving your editor.
A typical loop looks like this:
- generate_variants — produce a descriptive control plus one or more curiosity-gap subject lines
- launch_test — start the managed send and split your audience
- get_results — pull opens, clicks, and lift once data comes in
- recommend_next — get a suggested follow-up based on the winner
Writing curiosity-gap variants that hold up
The best curiosity gaps are honest — they tease a real payoff that the email actually delivers. When you have your AI agent generate variants, give it the body content and the offer so the gap is grounded rather than clickbait. This keeps your AI agent email marketing aligned with the message instead of optimizing only for the open.
Test one curiosity angle at a time. If you mix a question framing, a number tease, and an urgency hook into one variant, a win tells you nothing about which lever worked. Keep variants narrow so each subject line A/B test produces a reusable insight.
- Tie the tease to a payoff that exists in the email
- Isolate one curiosity mechanism per variant
- Watch click lift, not just open lift, before declaring a winner
- Reuse winning patterns through recommend_next
Getting started
marginal is hosted, so there is nothing to deploy — point your MCP client at https://marginal.sh/mcp, add your Bearer API key (marg_live_...), and start sending tools. The free tier covers 100 experiments per month, which is plenty to validate a curiosity gap email test across several campaigns.
Full tool schemas and client setup live in the docs at https://marginal.sh/docs/. Connect your agent once and every future subject line A/B test runs from the same place you write code.
- Endpoint: https://marginal.sh/mcp
- Auth: Bearer API key (marg_live_...)
- Works with Cursor, Claude Desktop, Claude Code, Windsurf, Cline, Continue, Zed, and OpenAI Codex
- Docs: https://marginal.sh/docs/
Get started with marginal
marginal is a hosted email marketing MCP server at marginal.sh. Sign up free, create an API key, and connect https://marginal.sh/mcp from Cursor, Codex, or Claude Desktop.
- 100 experiments/month on the free tier
- Four MCP tools: generate_variants, launch_test, get_results, recommend_next
- Listed in the MCP Registry as sh.marginal/mcp