Propagation of sound in shallow water is commonly analyzed in terms of trapped normal modes. The process of normal‐mode detection by adjustment of the response of individual hydrophones in a vertical array to match the pattern of a particular normal mode is normal‐mode filtering. Despite the nonorthogonality of the normal modes in the water column, normal‐mode filtering is based on the assumption that the modes are orthogonal. The error induced by this assumption is known as leakage. It has been estimated to range from 3% to 10% of a measured mode. In this paper, the error involved in treating the modes as orthogonal for a variety of bottom types, water depths, and source frequencies is analyzed, and a method for avoiding the error is demonstrated. Density and sound‐speed ratios are selected from models fitted to experimentally measured values. To render the results indicative of a broad range of water depths and frequencies, use is made of dimensionless variables. Leakages as large as 19% are calculated.