It is not known whether the variability of the glycemic index (GI) in different subjects is due to within- or between-individual variation. In addition, it is not known how large a difference in GI between different meals is clinically important for individuals with diabetes. Therefore, insulin-dependent (IDDM) and non-insulin-dependent (NIDDM) diabetic subjects tested four foods, with each food taken by each subject on two separate occasions. For each food, most of the variation of absolute glycemic responses was due to differences between the subjects. However, when the results were expressed as the GI, there were no significant differences between the subjects, and most of the variation was due to within-individual variation. Using the within-individual variance, we estimated the so-called "predictive difference" of GI values. Its reliability was assessed by consideration of published data from eight studies where different mixed meals were taken by the same group of subjects. There were 37 cases where the difference between the GI of any two meals was greater than the predictive difference. Of these 37 pairs of meals, the GI correctly ranked the glycemic responses in 36 (97%). We conclude that GI values for the same food do not vary significantly between different individuals. For a subject with NIDDM a difference in GI of 34 will predict the ranking of glycemic responses of two meals with 95% probability. The corresponding value for a subject with IDDM is 50.