I always assumed an average was performed over instantaneous velocity that fall in network segments. So, it would depend on the relative magnitude and the edge count on the graph in question you wanted to manipulate.
For a long very straight road represented as a single edge, I would assume if you had 100 actual cars, you'd get an average of the manipulated data rate and the real flow rate.
It works out to the same thing, more or less, but I'd guessed they used proximity nodes (junctions) and then when you reach another node use the time (discounting outliers).
Then do the calculations at the server for speed and such.
For a long very straight road represented as a single edge, I would assume if you had 100 actual cars, you'd get an average of the manipulated data rate and the real flow rate.