In summation experiments, responding to a compound stimulus is assessed after conditioning a response to each of its components. This simple experiment poses significant challenges to models of associative learning because of substantial variability in results. Here, I introduce a new method to quantify generalization from components to compound in summation experiments, which I apply to over 250 measurements of summation in rabbits, pigeons, rats, and humans. The analysis confirms that more summation occurs with stimuli from different rather than from the same sensory modality, although this is not the sole determinant of summation. A theoretical analysis shows that this finding is best accounted for by a model that includes both element sharing (Rescorla & Wagner, 1972) and element replacement (Brandon et al., 2000) in stimulus representations. I point out remaining gaps in our empirical and theoretical understanding of summation.