Synthetic Customers: Test the Funnel Before AI Traffic Finds It

A/B testing tells you what happened after real customers arrived. Synthetic customers offer something earlier: a structured way to expose likely friction before you spend the traffic.

What changed

Browser-capable agents can now encounter the actual interface rather than reading a screenshot or survey description. They can attempt checkout, react to missing information, compare prices, and document the step where a goal became impossible.

Where synthetic testing earns its keep

• Finding broken paths and ambiguous policies before launch.
• Generating edge cases across languages, devices, constraints, and goals.
• Ranking hypotheses before spending scarce real traffic.
• Stress-testing an agentic checkout from the machine customer’s perspective.

Where it lies to you

A model is trained to produce plausible behavior, not to reproduce your market. It may explain decisions more cleanly than real people make them, overreact to prompt wording, and miss social or emotional context. Validation must end with observed human behavior.

Synthetic customers are a wind tunnel, not the road.

A responsible operating loop

1. Ground personas in real segments and evidence.
2. Run the current and proposed experience with identical conditions.
3. Collect failure patterns, not fabricated conversion precision.
4. Turn patterns into hypotheses.
5. Validate the highest-value hypotheses with real users or experiments.

Emerging evidence is promising but early. SimAB reports 67% overall and 83% high-confidence accuracy across 47 historical A/B tests in its study; that is one research result, not a general guarantee. SimAB preprint, 2026.