There's an obvious, tempting way to sell this product: skip the slow, expensive part of user research and just ask the AI instead. We don't say that, and we're not going to, because it isn't true, and the gap between "directionally useful" and "a substitute for real customers" is exactly the gap that gets people into trouble.
What a synthetic panel actually is: a fast way to pressure-test a question before you spend real time and money finding out the answer the slow way. Would this pricing line confuse people? Does this onboarding flow explain itself? Is there an obvious objection we haven't thought of? A panel can surface the obvious version of that objection in minutes. It cannot tell you whether your actual customers, the ones who exist, will behave the way five generated personas predicted.
The distinction matters because it's checkable, not just a disclaimer we put in the footer. The diversity check on every panel measures something real and narrow: does this specific panel disagree with itself the way a real population would, or has it quietly collapsed into one opinion wearing different names. That's a check on the panel's internal validity. It is not, and can't be, a check on whether the panel's answers match reality outside the tool.
“No amount of internal diversity proves external accuracy.
A panel can be genuinely varied and still be wrong about what real people would say, because it's built from a language model's training data, not from your customers.
We see three ways this actually goes wrong when people forget the line. First, treating a synthetic panel's enthusiasm for an idea as validation instead of as one input among several: five generated personas liking your pricing page is not the same signal as five paying customers not churning. Second, using it to skip the step of talking to real users entirely, rather than to decide which three questions are worth their limited time. Third, quoting a panel's exact numbers ("73% would pay $15/month") as if they were survey data, when what actually happened is five personas each gave one qualitative answer and someone did the math on five data points.
The honest use case is upstream of all of that: a cheap way to find out what you don't know yet. Run the panel before you write the real survey, not instead of it. If four personas raise the same objection independently, that's worth taking into the room with real users, framed as a hypothesis, not an answer. If the panel is unanimous, that's often more suspicious than reassuring: real audiences rarely agree that cleanly, and the methodology page exists partly to help you notice when your panel is being suspiciously agreeable instead of actually diverse.
So the line we won't cross:
“No claim on this site says a synthetic panel knows what your customers think.
Every claim we do make is about something checkable inside the tool itself: whether this panel disagrees with itself, whether its age spread looks like a real population's, whether its answers to your specific question are diverse or collapsed. That's a smaller, less exciting claim than "replace your research team." It's also the only one we can actually stand behind.