Actually, you can argue with your experiment, because you did it wrong.
To explain your problem: IF the main man is guilty (1/3 of the time), then only one of the other guys is guilty, and the guard HAS to point to him. He will point at him 100% of the time. So 1/3 of the time this situation occurs, and the not-pointed-to guy gets off free. Your experiment, however, does not account for the fact that the guilty man must always be chosen: your experiment has the guard choosing among the other two ~50/50 ALL the time.
Let us analyse that. This case, which obviously occurs 1/3 of all times (and as is borne out by your experiment with no removals), leads to the not-pointed-to, not-main guy being innocent. The main guy is, by assumption, guilty.
IF the main man is not guilty (2/3 of the time), then the other guys are BOTH guilty. The guard can point at either of the other men. Now the odds are 50/50 which of these two will be selected.
Let us analyse this. This case, which occurs the other 2/3 of all times, leads to the not-pointed-to, not-main-guy as being guilty. The main guy is, by assumption, not guilty.
THEREFORE, the not-pointed-to guy is guilty 1/3 of the time, the pointed-to guy is guilty ALL the time, and the main man is guilty 2/3 of the time.
Or think of it like this, in a much less-rigorous way: all three have a 2/3 chance of being guilty to start with. When the guy is pointed-to, he gets an additional 1/3 chance of being guilty. Somebody has to donate that 1/3 chance, and that has the other guy, since the guard was NEVER going to point at you so you weren't involved in that selection.
Or simply: If main guy is guilty, then the guy not pointed to is always innocent. Therefore, if you are NOT the main guy AND the guard doesn't point at you, you've increased the chances that YOU are innocent -- if you were guilty, he'd point at you half the time, if you were not guilty, he'd NEVER point at you, and combine half with never and you get something LESS than half.
A more analogous experiment that you could try is to pick the innocent guy out of the hat FIRST, then remove one of the guilty, non-main guys, then see how often the main guy was innocent and how often the not-pointed-to not-main guy was innocent. After all, in the example, the guard knew which one was innocent even before the main guy asked.
Even better, if you know anything about computer programming, you can make a simple program to convince you. The program randomly determines who is guilty, and you play the role of the main guy. You hit enter, the computer tells you (choosing at random if necessary) one of the other guys who is guilty, then you hit enter again and see who is actually guilty. There's one already made here:
http://chrisc.freeshell.org/random/pages/montyhall.html
ALTHOUGH given your bias against random webpages it could just be a conspiracy to give you the wrong answer and not ACTUALLY an instantiation of the code shown on that page.
-----
Before ignoring a website as "random" you might want to read what it says -- it does not matter how credible that site is, if the reasoning is sound.
But maybe you trust wikipedia more?
http://en.wikipedia.org/wiki/Monty_Hall_problem
Or a math site?
http://mathforum.org/dr.math/faq/faq.monty.hall.html
http://mathworld.wolfram.com/MontyHallProblem.html
Or the site whose very URL refers to the problem in question?
http://www.montyhallproblem.com/
Honestly, I'm not sure what kind of website you were expecting. The CIA?
In general, don't believe everything you see on the Internet, but also don't
disbelieve everything just because it's on the Internet, at least not without a second look when a majority of people are telling you you are wrong.