@pezus
You seem to answer your own question. If the games eventually disappear from the shelves completely then logic tells us that the game has sold through at least its first production run. So if this site is telling us it sold closer to a million relatively soon after its launch. Then that should mean that the initial production run was a million units, and if it did well enough early on then a second smaller production run probably happened. It is simply a matter of basic deductive logic.
We all know how the game is played. Games rarely get destroyed. They just get discounted until they all get sold. The only exceptions being faulty games, or games that just can't be sold. Any game that sells over half a million is without question good enough to not fall in the latter category. So like it or not even if this site doesn't say so. It is highly likely that those games sold a million or more copies.
On the subject of this sites tracking. I assume that the majority of what we see are generic extrapolations, and you can probably safely assume that they will never admit to that. After a game leaves the ranks of the top sellers they just take the previous weeks figures, and plug them into a standardized mathematical progression. Which will carry the sales down to a point where they don't even bother to pretend that they are actively tracking the sales at all. It really is like clockwork.
When this site provided weekly figures on the games profile page you could do the math, and see that almost every single game followed the same predictable progression, and that is purely nonsense right there. Buying patterns are a product of human behavior, and there is nothing less predictable in nature. Even games three or four months out from their launch should have detectable spikes, or even inexplicable dead patches. They shouldn't fit perfectly into any kind of curve.
Really a lot of it isn't even educated guess work. It is just the same rule of thumb applied to all. Unless what is being tracked has what we all used to call legs. Anyway what this site provides is decent short term tracking based upon a limited data set and secondary sources. It isn't a question of accuracy so much as it is a question of respectable data not being present. You shouldn't be drawing grand conclusions at all. Nobody should for that matter.
P.S.S
Actually it isn't better then nothing. Bad data can lead you to draw the wrong conclusions. If you want to use the data then limit yourself to the most reliable part of the data sets. Namely what the market is doing in the here and now. Not sales where this site could be off by hundreds of thousands of units.