As a rough algorithm, I'd guess that counting each distribution (regardless of revision) once per IP per day would be a reasonable first-approximation.

Part of me agrees that revisions should probably be ignored, but then the counts will likely skew upwards if there are many revisions. However that skew could be viewed an indicator that the modules development is being followed regularly and not just downloaded as a curiosity. It's a difficult one.

As for the "once per IP" idea, I have found with log analysis that IPs can be deceiving because of proxy servers. So I think that we should be wary of weighting the IPs too much.