Like c4c said, how do you know you aren't being cheated with game benchmarking too? In most cases I'd assume it's easy for driver designer teams to detect whenever a game is running a timedemo and easily cheat in that. It's way hard to detect such cheats and rarely do people do in-game benchmarking so we never get to backup the findings that we see from the benchmarking.
Give me two weeks, and I'll tell you how I know. Working on an article about that right now, and I have a solution. It ain't pretty (yet, this is sort of a bastardized thing I've done, but if other people take the idea and run with it, it'll be great), but it works, and it gives reliable scores.