This actually came up in a discussion I was having on another forum. After the argument started, we were fortunate to have one of the developers of a particular road dyno package join the discussion. It quickly devolved into higher math, but the summary of the situation is that the math used to make the plot has a pretty serious problem. Interestingly, this is also true of the lower-end physical dynos.

The root of the problem is that as the dyno or dyno software sample speed (of the engine [RPM] or of the rolling drum [RPM]) over time, reality gets in the way and the change in RPM over time isn't very consistent from sample to sample. Sometimes, the RPM may even drop between samples (leading to NEGATIVE power). To deal with this, the software (road dyno or real), applies smoothing to the graph in the form of fitting a high-order (x^16) polynomial function to the data.

*Most* of the time, that works out just fine. However, from time to time, the function will have an inflection point (change in curvature) that the real power curve doesn't. Obviously, this is a BIG problem for tuning.

Here's an example from my RS. See that hook at the top end? That's completely fake, not there at all. It's totally an artifact of the math.

The safest/easiest to get plots that don't have math weirdness are RPM vs time (a line running generally up and to the right, getting higher, faster means more power), or a plot of dRPM/dt vs. RPM (higher is better). Both of those can easily be plotted in Excel and would allow you to do overlays to compare different runs.