We assessed the accuracy of scat-sampling methods in relation to sources of bias (statistical independence of the data and definition of the sampling unit) and precision (sample size). We developed a method to quantify diets of predators accurately in a study of diet selection by wolves (Canis lupus) during 3 winter seasons (1999–2002) in the Western Alps. The best sampling design to avoid pseudoreplication was the “additive method,” where the presence of a carcass, estimated by either a collection of scats or a carcass itself along the travel route of a wolf, was considered 1 sampling unit. Although roe deer (Capreolus capreolus) were the primary prey used by wolves in the area, red deer (Cervus elaphus), recently reintroduced prey present at low density, were selected in winter 2001. We evaluated the optimal sample size for a given question using Monte Carlo simulations. At small sample sizes, slight increases in sample sizes caused large reductions in the standard error, greatly improving the precision of the estimates of percentage of items in the diet. Estimating the number of rare prey species used by wolves, such as red deer in our case study (<2% of the diet estimates), was possible if the minimum sample size was greater than 10–40% of the population of carcasses. We emphasized the importance of the additive method to improve the accuracy of estimates of diet selection by carnivores.