Computational studies of transcription factor binding to DNA are usually 
based on a simple matrix model (PWM) of sequence-dependent binding energy. 
For some transcription factors, the best matrix we can construct is alarmingly
non-specific, predicting many binding sites that are not known to be 
functional. If these sites are in fact spurious, the validity of the matrix
model as a physical description of binding specificity is called into
question. This is unfortunate, since the model is a very convenient framework
for physical thinking about gene regulatory networks. This has motivated us
to develop a species comparison approach to assessing the functionality of
populations of such sites: we compare putative sites with aligned sequences
in other genomes and ask whether the probability of mutation depends on
position within the site. We find a very distinct `footprint' in the mutation
probability which has the feature that mutation is most strongly suppressed
where it would have the biggest effect on the binding energy as assessed by
the matrix model. This suggests that many of the apparently spurious sites
are functional and also that something related to the matrix model estimate
of site binding energy is what is conserved between species. Apart from
theoretical reassurance, this analysis casts up new binding sites and new
regulated genes which may be of interest to experiment.