When "best" is interpreted as "minimum variance unbiased estimator," then the answer is cokriging. This technique views the three datasets as realizations of three correlated spatial stochastic processes. It characterizes the process and their correlations by means of a preliminary analysis, "variography." The variography can be informed by additional considerations such as the relative quality of the datasets: this would cause the analyst to prefer variogram models that assign higher variances to the lower-quality datasets. Then, the co-kriging algorithm itself uses the data and the variographic output (a set of "pseudo co-variograms") to predict any desired linear combination of the three datasets (such as the mean values of dataset "A" within specified grid cells).
A simple, limiting version of co-kriging supposes the three datasets are not correlated with each other at all. In this case, each of the datasets can be kriged separately. (This means a variogram model is fit to each dataset separately and used separately to interpolate the data values onto a grid of specified locations.) The kriging output includes a kriging variance. Assuming each dataset represents the same physical quantity, the question becomes how to combine the three estimated values at each grid point into the "best" possible estimate of that quantity. The answer is to take a weighted average of the three estimates, using the reciprocals of the kriging variances as the weights. The result remains unbiased because the three inputs are unbiased by construction; weighting inversely by variance assures the result has the smallest variance, which is exactly what "best" was assumed to mean.
Kriging and co-kriging are available in the Geostatistical Analyst add-on for ArcGIS (at an extra cost) and in freely available software such as the gstat package for R. These are not activities to be undertaken casually: they are sophisticated, comprehensive analyses of the data that become valid only as a result of reasoned, accurate statistical characterization. Although software for kriging and co-kriging has long been available in many GIS environments, IMHO kriging is not an activity to be performed solely by a GIS analyst, because it requires close collaboration both with a seasoned statistician and an expert in the field of study. (Occasionally two of these three roles or even all three may be adequately filled by the same person, but this is rare.) It is one of those especially dangerous capabilities of a GIS because kriging output is easily created by anyone who can push the right buttons (a task that can be learned in 30 minutes with ESRI's excellent tutorial on Geostatistical Analyst, for instance) and will often look correct but is just the "GO" part of the proverbial GIGO.