A proposed solution, for the critique of any mathematicians here:
Always split each channel into ten 10-percentile bands, instead of just taking an overall average. Then model total expected revenue per channel based on these bands.
Perhaps as few as three 33-percentile or five 20-percentile bands would be enough if the number of samples is low when first testing a channel.