Four surface‐wave tomographic models in the Pacific Northwest and a combined CRUST2.0 and AK135 model are tested and validated systematically. Synthetic Green’s functions calculated with the models using a finite‐difference method are compared with empirical Green’s functions at periods of 7–50 s. To ensure high‐quality signals, empirical Green’s functions are extracted from the ambient noise cross correlation of vertical‐to‐vertical components between station pairs that have up to a decade of recorded data. The observed and synthetic Green’s functions are cross correlated at multiple frequency bands to determine phase delay times and cross‐correlation coefficients. The delay time predicted by the CRUST2.0 and AK135 model is predominantly positive and is linearly dependent on interstation distance, indicating that the combined model is, on average, too fast for the Pacific Northwest. Among the four shear‐wave velocity models, CUB and one model derived from regional tomography exhibit moderately and weakly negative linear trends, respectively, between the delay time and interstation distance, a result indicative of a slower‐than‐actual velocity. The delay times of the other two models are normally distributed with an approximately zero mean and without any apparent relationship with interstation distance. The cross‐correlation coefficients are more scattered at short periods, reflecting unresolved heterogeneities of the crust structure in these models. The misfit between the empirical Green’s functions and synthetic waveforms suggests the need for a better‐resolved crust and uppermost mantle velocity model, which is critical for the precise estimate of ground motion for seismic hazard evaluation and understanding of the tectonic processes of the Pacific Northwest.