Individual observations are routinely used in livestock evaluations. In some cases, pooled data representing the joint but not individual performance of a group of animals may be available. For example, pooled feed intake may be measured on a pen of livestock. The usual mixed model approach to genetic evaluation can still be applied as an exact method in this setting, provided incidence and residual variance-covariance matrices are suitably modified to account for the pooling. Approximate evaluations may be achieved by treating average performance as if it pertained to each individual in the pool. Theoretical accuracies can be obtained as a function of elements of the inverse coefficient matrix. A 3-generation data set representing 1,000 animals with feed intake observations from 49 sires and 200 maternal grand sires was simulated with heritability of 0.34. Individual records were pooled to represent circumstances in which animals with records were collectively measured. Animals were allocated into pens at random, by sire, or by maternal grand sire. Simulation was replicated with unique fixed effects for each pen. Following evaluation from each method, the empirical accuracy or product-moment correlation between true (simulated) and estimated merit could be quantified. The analysis of individual observations resulted in empirical accuracy of 0.63 for animals on test and 0.77 for their sires. Pooling the observations in pens of 2, 4, or 12 animals reduced empirical accuracies for animals on test to 0.50, 0.41, and 0.21 when pooling was at random and 0.53, 0.47, and 0.34 when pooling was by sire. Simulating a fixed pen effect representing 10% phenotypic variation, but ignoring that effect in the evaluation minimally reduced empirical accuracies to 0.52, 0.46, and 0.33 when pooling by sire. Theoretical accuracies were in close agreement with empirical accuracies when the exact method was used. The approximate method that treated averages of pooled data as if they were individually observed overstated accuracy and should not be used. Selection on the basis of pooled observations can be almost as effective as using individual observations when pool sizes are small. The exact method to account for pooled data is no more complex than conventional procedures.