Most subduction zones have associated long‐wavelength anomalies in the lithospheric magnetic field observed at satellite altitude. We model the 13 subduction zones defined by seismicity and seismic tomography using vertically integrated magnetizations that are increasing, level, or decreasing away from the trench. These mimic end members of a magnetized mantle wedge, a uniform layer, and a magnetized dipping lithospheric slab. They are added to a global model of vertically integrated magnetization based on continental and oceanic geology. We find the dipping slab places the anomaly too close to the trench, while the other two fit the data equally well and use the level model in the main part of the study. Anomalies at the Sunda, Aleutians, Cascadia, Central American, and Kamchatka‐Japan zones are well modeled by uniform magnetization of differing susceptibilities and spatial extents. We show the South American anomaly is weak because the magnetization lies mainly in the null space that produces no external potential magnetic field. There is no anomaly associated with the Ryukyu system, possibly because the present subduction started too recently for magnetization to have formed. The magnetic anomaly stretching down the Baja California peninsula is not present in the prediction because there is no seismicity on which to base a slab geometry, but recent tomography suggests a fossil slab there and we propose historic subduction as the origin of the Baja magnetic anomaly. Finally, we discuss the mineralogical origins of the magnetization and favor serpentinization of the region above the subducted plate.