Is there a name for the fallacy of ignoring marginal effects at the tail end of a probability distribution? I see it here incredibly frequently.

There will almost certainly be some number of people who would have stopped by Amazon right now and made some impulse purchases. At the scale Amazon operates, the increase in inconvenience to push off the marginal purchase as a function of inconvenience is almost certainly miniscule (See frequent reports on how milliseconds of page load time affect the likelihood of purchase)

Exactly. People say they lose $X/hour. The next hour when they come up, is the order rate back to $X/hour, or $X + delayed demand / hour?

(Surely there's some loss from being down, but it's not a simple loss = current order rate * downtime argument).

I'd be more concerned about the PPC ads leading to a 500 error.

Not completely zero - I can easily imagine how someone trying to buy a gift during a fifteen-minute break would go to barnesandnoble.com or whatever.

I can attest to this being the case. I mentioned it elsewhere in the thread, but we're in e-commerce and when walmart.com went down around black friday last year we saw a 20megabit jump in traffic until their site came back up... and we're only one e-com provider out of many.

Amazon might be able to measure this effect, by taking comparing to their projections over the next day. Imperfect, but...

... or go to costco.com, which is a great site that oftentimes can beat amazon's price.

Not if that customer moved on to a competitor’s site and made the purchase there.