The new technology of protein binding microarrays (PBMs) allows simultaneous measurement of the binding intensities of a transcription factor to tens of thousands of synthetic double-stranded DNA probes, covering all possible 10-mers. A key computational challenge is inferring the binding motif from these data. We present a systematic comparison of four methods developed specifically for reconstructing a binding site motif represented as a positional weight matrix from PBM data. The reconstructed motifs were evaluated in terms of three criteria: concordance with reference motifs from the literature and ability to predict in vivo and in vitro bindings. The evaluation encompassed over 200 transcription factors and some 300 assays. The results show a tradeoff between how the methods perform according to the different criteria, and a dichotomy of method types. Algorithms that construct motifs with low information content predict PBM probe ranking more faithfully, while methods that produce highly informative motifs match reference motifs better. Interestingly, in predicting high-affinity binding, all methods give far poorer results for in vivo assays compared to in vitro assays.