If a player is hovering around his "real rating," win percentages tend to be fairly similar over a long series of games - around 50% - regardless of rating. A highly skilled player will generally accumulate more wins than losses before the two start to even out. Consider a 2000 player. If he gains 20 points per game, he'd have to win 20 games more than he loses to get to 2000 - if he's played 40 games, then his win percentage will be about 75%. If 2000 is his actual level, and he keep playing, his win percentage will creep ever closer to 50%, but never quite get there, barring a statistical anomaly wherein he plays mostly people rated above him.
Yale explained well why win percentages aren't always significant - they can be an indicator, but the nature of how overpowered civs work means that the win percentage won't be anywhere near being conclusive evidence. Since Set rooks tend to gain higher ratings than their skill, they hit an equilibrium at a higher rating level than they should be at, where they win about 50% (when in reality their skill should let them win perhaps 30% or less against players of the rating level they play at).
For a good example of your quintessential Set rook and lamer, look here: http://aom.holbert.org/stats.php?user=slav_knight&matchType=ZS_Custom&timePeriod=ZS_AllTi me&query=st&tz=0 This guy's actual skill is around 1800 to 1850 at the most, but his rating exceeds that (note that he also stopped playing once he hit a "nice" number, and one that allowed him to join a few 1900+ games before being found out as a dude with a bloated rating and bad personality to boot).Now, as for the combined brain trust of R&A (Rust and Age), and whether or not Set is overpowered at lower rating levels, let's examine some assumptions.
1. The lower the rating level, the more random the outcome. Lower rated players are less able to exploit their civ, their map, their position, generally less knowledgeable, generally slower, and so on - in short, the outcome of games tend to be more random. This alleviates imbalances somewhat. To take it to an extreme, think of how a game between two deaf and blind players would turn out.
2. Conversely, the higher the rating level, the more everything matters - civ, map, position, all of these things tend to get more important for the outcome of a game the higher the skill level of the involved players.
3. When something is too good, it is thus first discovered at the higher skill levels, and then spreads downwards. It tends to spread relatively quickly through the upper echelons, then slow down as it reaches the lower ones. People who play at a higher level tends to stay better informed of what is the 'current strat to copycat' and also understand better why that is so. Understanding of why something works may be lost on lesser players.
4. Something that is overpowered is overpowered on all skill levels. However, the lower the skill level, the less the ability to exploit something that's overpowered, and the less overpowered it seems to the rookie player. The ability to exploit game imbalances takes more time to develop at lower levels than higher ones, some imbalances more than others. Once this has been learned at a certain level, however, it will increasingly dominate that level of play.
5. A player that extensively abuses an imbalance will generally win games he shouldn't, and play at a higher rating level than his skill indicates. If you are a true 1800 player up against someone that has only gotten to 1800 through exploiting an imbalance, your chance of winning should still lie around 50%. This applies regardless of rating level, but is increasingly true at lower levels, where the outcome of games is more random.
6. Major changes in strategy paradigms take longer to spread downwards through the ratings. For an example of this, look at the flush in AoK. It got popular at a high level fairly early, but it took a very long time before it reached the average players. You'd find rooks sticking to castle rush build orders they'd copycatted and honed for months long after it was established that flushing was a superior strategy. The same can be said to be true of AoM at this point, though in a different way - Set has been found to be superior to the standard Norse RC strategies; however, this change is taking significantly more time to spread downwards than strategies that were relatively easier to execute. It is spreading downwards, however, and it will hit pretty much all levels of rated eventually if not patched or changed. Arguing that 'it only affects experts' is a foolish argument at best, and seems born more out of a form of willful pride and refusal to accept that one is wrong than actual rational thought.