xcorr returns a 2*M-1 cross-correlation sequence for a vector length M from lags [-M:M] w/ the zeroth-lag in the middle of the returned vector. The auto-correlation (0-lag) value is going to be the max() so you need to restrict the search to lags>=1 and as well look for the maximum away from the initial decaying self-correlation around zero that can be significant depending on the shape of the two signals.
And, of course, there's the "minor" detail that the maximum correlation may not reflect the desired lag if the two signals aren't particularly self-similar with a phase lag only the difference as opposed to some large feature in one or the other.
