A scheme is described whereby the error associated with the least well-resolved model eigenparameter in a magnetotelluric survey is reduced by focusing data collection on a specific range of frequencies. The scheme also gives a quantitative estimate of the statistical error associated with the least well-resolved model parameter, and thus provides an objective criterion to the operator regarding when to cease data collection at that location.The scheme is based on a linearization of the relationship between variations in the model parameters and the changes thereby introduced to the computed response function. The matrix of partial derivatives describing this linearization is factored orthogonally by a singular value decomposition.The scheme is illustrated by applying it to a synthetic data set. Also, the algorithm has been coded in Basic on an HP9845 and employed in the field. An example is given of its field operation in a sedimentary basin environment.