Using iterative experiments to uncover causal links between critical policy details and outcomes helps to optimize policy design. This paper studies a large-scale staged fiscal stimulus program conducted during the COVID-19 pandemic, in which a provincial government in China disbursed digital coupons to 8.4 million individual accounts in consecutive waves and updated the program design each time. We find that ruling out unproductive program features leads to a pattern of increasing treatment effects over the waves and that program design matters more than the size of the fiscal stimulus in boosting spending. Our results show that (i) general coupons with no constraints on where the vouchers can be redeemed are more effective than specialized coupons in stimulating consumption in the targeted sectors; (ii) coupon packets with fewer denominations and shorter redemption windows tend to be more effective; and (iii) low-income residents and non-local residents are equally or even more responsive to the coupon program than other groups. Our results illustrate that generating variations in iterative policy experiments, combined with a timely assessment of individuals’ responses to marginal incentives, optimizes program design.