In this paper we study the effect of the sampling strategy on the achievable accuracy in linear system identification experiments. We consider the experiment as being composed of a number of subexperiments and we impose the constraint that the sampling rate should be constant in each of the subexperiments. We investigate the effect of having different sampling rates in each of the subexperiments and we show that the optimal information return can be achieved by use of a finite number of sub-experiments. We develop the theory in two parts : the first relating to the identification of a stochastic process from output observations ; the second relating to the identification of an input-output transfer function from noisy observations. We present a number of examples which show that the appropriate choice of sampling strategy can be of paramount importance in identification experiments. We also show that the design can be extended in a straightforward manner to safeguard against a single non-informative experiment in the case of diffuse prior distribution for the parameters.