Do what sounds good before what was accurate. As Lee (and one of his favorite quotes) point out, it's an illusion. You can make that illusion sound approximately what it would have been like, or you can do something different. At the end of the day, most people will grade the end result on how it sounds, not if it was accurate, although there are some glaring inaccuracies that one should avoid. For example, avoid moving musicians around (so if you have 5 people L to R on stage, don't flip 1 and 4), especially if you have an introduction where people say "and to my left is Johnny Two Shoes McGee and to his left is Dingly Dolittle" and they start playing only to be on a different spots in the mix than the vocalist just said. If you really want to move him/them, then eliminate that speech or other traces. This is easier said than done so I try not to do it. Second reason to adhere to a rough approximation is due to the time delay of sounds reaching other mics. For example, if I have a semi-circle, sounds originating from #2 will be heard in bleed sooner and more prominently in 1 and 3 than in 4 or 5. Avoid altering that balance too much. All of this can be minimized via two avenues; first off, use the polar pattern to your benefit. If you have someone in location 3, but you don't want 2 or 4 in that channel, then consider using a hyper which has better side rejection than a card which has excellent rear rejection. Likewise, for 1 and 5, think about using the cards as then you can put that null pointing back across the stage to eliminate as much of that as possible. You'll never get rid of all of the bleed, but that's ok (it helps with a natural reverb, and it's difficult to coherently hear the bleed when you have lots of other sounds in your mix that are at least +15 or 20db louder).
These are guidelines. Some times I've tossed them out the window because I had tiers (so think about 8 mics across 3 rows, the middle rows have both the back row content and the middle row, and the front row has the middle row and the front row). That sort of thing requires a lot of forethought to really do well, and you may or may not have that time to think about it when you're throwing stuff on stage. So the short of it is, if you have the channels and equipment (and you are short on time or planning), spot it. If you have a lot of time to plan and test, then you can get fancy with planning bleed and stuff like that. This is all genre sensitive as well; jazz and classical do better with bleed (and classical almost mandates it) than rock, country or electronica, but it's a sliding scale, not a rule. Also, when in doubt, trend toward mono. There are other tricks you can do to create "stereo spaciousness" without it being flat, even though the same content is basically in both channels.
As Lee also notes, headphones have a different stereo representation than speakers. I'll add that the difference is greatest in panning content over time-delay content (which sounds different but is less likely to move around or get louder than pan/scan).
I'm at work, so I hope my rushing didn't cause this to be too complicated, others covered lots of it so I tried to just add different info. Let me know if there is something I've said that doesn't make sense or needs more explanation.