Mirror Panel Inc. Information

Billion-dollar decisions still get made by eight people in a room. Focus groups are slow, expensive, and impossible to run twice — which means they can never tell you how many people actually agree. Most of the industry has responded by trying to replace them. We think focus groups need more horsepower to deliver what they always promised. Mirror Panel builds virtual people. Two thousand of them, with demographics sampled from the US Census before any AI model is called — balanced across age, region, race and ethnicity, household income, education, urbanicity, gender, and household. You can poll them, interview any one of them and ask why they answered that way, and come back to the same panel next month. How it works. Ten ideas go into Mirror; two come out ranked, with a measured spread on each. You test those two with a live focus group — real humans, real hesitation, the objection nobody anticipated. Their verdict returns as calibration, and the finding projects across two thousand census-weighted respondents. The room stops being your sample. It becomes your calibration. That last step is grounded in peer-reviewed work: exposing frontier models to a human benchmark improves their accuracy by 17–28% and tightens their prediction intervals (Schoenegger et al., Science Advances, 2024). Why measurement matters. Swap the AI model behind a study and the answer can move by up to 38 points. We measured it: 98% of the variation in a result is which model you picked — a hidden bias every single-model tool ships blind. Mirror runs an ensemble and quantifies the disagreement, so the ideas that survive are the ones that hold across models rather than the ones a single model happened to favor. We benchmark against real human studies — Morning Consult brand trust, GroceryTV in-store sentiment, YouGov apparel consideration — with hypotheses and pass thresholds fixed before every run. Founded in New York.