Silencing specificity is a critical issue in the therapeutic applications of siRNA, particularly in the treatment of single nucleotide polymorphism (SNP) diseases where discrimination against single nucleotide variation is demanded. However, no generally applicable guidelines are available for the design of such allele-specific siRNAs. In this paper, the issue was approached by using a reporter-based assay. With a panel of 20 siRNAs and 240 variously mismatched target reporters, we first demonstrated that the mismatches were discriminated in a position-dependent order, which was however independent of their sequence contexts using position 4th, 12th and 17th as examples. A general model was further built for mismatch discrimination at all positions using 230 additional reporter constructs specifically designed to contain mismatches distributed evenly along the target regions of different siRNAs. This model was successfully employed to design allele-specific siRNAs targeting disease-causing mutations of PIK3CA gene at two SNP sites. Furthermore, conformational distortion of siRNA-target duplex was observed to correlate with the compromise of gene silencing. In summary, these findings could dramatically simplify the design of allele-specific siRNAs and might also provide guide to increase the specificity of therapeutic siRNAs.