Coastal marshes record a 6500 yr history of coseismic displacements in great earthquakes at the Cascadia subduction zone. We compiled estimates of coseismic displacement for past megathrust events based on correlations with megathrust-triggered turbidites, and estimated megathrust slip based on comparisons of marsh displacements with dislocation model predictions. Age-correlated marsh data are compatible with event rupture extents defined by the published turbidite record, and a 6500 yr mean recurrence interval that increases northward from ∼230 to ∼480 yr. Within the constraints of the marsh data, the width of the coseismic rupture zone generally agrees with the downdip width of the interseismic locked zone inferred from geodetic and thermal data. In southernmost Cascadia, where the model does not include the complex deformation near the Mendocino triple junction, the coastal data may be better fit by a model with an ∼25% narrower rupture than that inferred from regional geophysical data. At each coastal marsh site, coseismic displacements are roughly similar from event to event, independent of the time since the previous event. Slip in the A.D. 1700 earthquake was consistent with the preceding interval of strain accumulation (∼200 yr) only at the northern and southern ends of the margin, but it was apparently much higher in southern Washington and northern Oregon, possibly indicating postseismic contamination and/or catch-up coseismic slip to make up for a deficit in the preceding event. Overall agreement between the dislocation models and the marsh data for most of the margin implies that such models can be usefully applied to rupture and ground shaking predictions.