A synthetic reference (SR) or fiducial mark is a target that has been designed and placed within a scene in order to help to locate objects. When a scene containing a small SR is acquired through a pixelated device (such as a CCD camera) and this SR is located using a digitally computed normalized correlation search, there is a lowering of the correlation peak if the target SR is displaced a nonwhole number of pixels with respect to the position of the previously acquired model SR. We propose a framework where the lowering is related to detection reliability, in order to compare the performance of different sizes and shapes of SR. This framework has been applied to eight SRs having simple shapes, using synthetic scenes. Results show large differences in their detection reliability according to their shape. A SR in the shape of a 45-deg-tilted square shows the best behavior among all the studied shapes, and one can take advantage of this fact when designing an application dealing with small-size SRs.