Evaluate the model in log-log space, transform those values back to non-log'ed values and then compute the error and total sums of squares from those values. R-square from the fitted model is the goodness of fit in the transformed space which is inflated owing to the contraction in magnitude of the log transform. But, if it isn't pretty high there, it will be turn out to be low in "real" dimension.
ADDENDUM
I do presume you're aware the definition of
R-sq=1-SSres/SStot
where SSres,tot is the residual and total sum of squares, respectively. And those are easily obtained in Matlab via dot
