(1) what do you base the variation on? 

Soil type, rainfall estimates (loss) x soil type, to some degree expected yield (will the sand make good corn this year?)

(2) how big of a range do you vary? 100 units to 250 units, etc? 

Based purely off of UW N guidelines - the best data I know of.   I am trying to incorporate a site-specific N "deficit" calculator - based on Cornell's publications which is now being resold through Climate Corp.  Total variance is not much, but early N can vary ~ 30 units (heavy ground gets higher rates) and later applications fairly narrow (proportionately more N on the sand ground on late applications) 

(3) and the big kicker here - how do you measure how close you are to being correct? I don't mean how do you think you are correct, how ARE you measuring your correctness? 

Cannot  be done without a blinded trial - no software can do that.  A rough estimate is end of season soil nitrate and stalk nitrate testing - looking for over application.   Remote sensing should - someday - add additional guidance.