We study when and how randomization can help improve the seller's revenue in the sequential screening setting. Using a model with discrete ex ante types and a continuum of ex post valuations, we demonstrate why the standard approach based on solving a relaxed problem that keeps only local downward incentive compatibility constraints often fails and show how randomization is needed to realize the full potential of sequential screening. Under a strengthening of first-order stochastic dominance ordering on the valuation distribution functions of ex ante types, we introduce and solve a modified relaxed problem by retaining all local incentive compatibility constraints, provide necessary and sufficient conditions for optimal mechanisms to be stochastic, and characterize optimal stochastic contracts. Our analysis mostly focuses on the case of three ex ante types, but our methodology of solving the modified problem can be extended to any finite number of ex ante types.