As an alternative to spectral methods, stochastic self-similar slip can be produced through a composite source model by placing a power-law scaling size-frequency distribution of circular slip dislocations on a fault surface. However these model do not accurately account for observed surface rupture behaviour. We propose a modification to the composite source model that corrects this issue. The advantage of this technique is that it accommodates the use of fractal slip distributions on non-planar fault surfaces. However to mimic a surface rupture using this technique, releasing the boundary condition at the top of the fault, we observed a systematic decrease in slip at shallow depths. We propose a new strategy whereby the surface is treated like a reflector with the slip being folded back onto the fault. Two different techniques based on this principal are presented: the first is the method of images. It requires a small change to pre-existing codes and works for planar faults. The second involves the use of a multi-stage trilateration technique. It is applied to non-planar faults described by an unstructured mesh. The reflected slip calculated using the two techniques is near identical on a planar fault, suggesting they are equivalent. Applying this correction, where reflected slip is accounted for in the composite source model, the lack of slip at shallow depths is not observed any more and there is no systematic trend with depth. However, there are other parameters which may affect the spatial distribution of slip across the fault plane. For example, the type of probability density function used in the placement of the subevent is also important. In the case where the location of maximum slip is known to a first order, a Gaussian may be appropriate to describe the probability function. For hazard assessment studies a uniform probability density function is more suitable as it provides no underlying systematic spatial trend.