Well, that could make sense. On a log scale, much of your data could be either very far away from the central portion of the 3 Gaussian lobes or clustered very close together, and therefore don't provide very informative measurements. So, use a linear scale to sample the curves. Or, sample on the log scale more finely and apply weights to your objective function, so that the samples near the central lobes get more priority.
Also, fminsearch is probably not a good tool to use with the number of unknown variables that you have. You should probably use lsqcurvefit instead, ideally with SpecifyObjectiveGradient=true..