High-density oligonucleotide arrays are among the most rapidly expanding technologies in biology today. In the GeneChip system, the reconstruction of the sample mRNA concentrations depends upon the differential signal generated by hybridizing the RNA to two nearly identical templates: a perfect match probe (PM) containing the exact biological sequence; and a single mismatch (MM) differing from the PM by a single base substitution. It has been observed that a large fraction of MMs repetitively bind targets better than the PMs, against the obvious expectation of sequence specificity. We examine this problem via statistical analysis of a large set of microarray experiments. We classify the probes according to their signal to noise (S/N) ratio, defined as the eccentricity of a (PM,MM) pair's "trajectory" across many experiments. Of those probes having large S/N (>3) only a fraction behave consistently with the commonly assumed hybridization model. Our results imply that the physics of DNA hybridization in microarrays is more complex than expected, and suggest estimators for the target RNA concentration.