The trick is that that it's no longer a 50-50 or 33-66 scenario when it's broken into two stages. Of course, if I simply present you with two doors and say one has a car and the other a goat, it will be a 50-50 and wont matter whether you then switch after your initial choice. Similarly, if I show you 3, open none of them, then ask you again, your odds will still be 33% of getting the door with the car, regardless of whether or not you switch. The difference in this scenario is this: on your first guess, each door has a 1/3 chance of having the car. Say you pick door #1. When the host essentially looks at #2 and #3 and tells you "door number two has a goat behind it," the 2/3 probability between doors #2 and #3 has now shifted entirely to door number 3, since door 2 has been eliminated. It's a filtering process that only applies to the two doors you havent chosen, telling you nothing about your original choice but improving the odds of the one that remains of the two that were filtered.