Just read a bunch of reviews, try and discard ridiculous results as even experienced benchers and unbiased reviewers will not spot a mistake. Sometimes a game just won't run right, installed wrong, did something stupid and runs strangely. On I believe 470gtx sli type review at Toms one of the scores dragging down AMD's xfire scaling scores were Crysis basically running horrifically in xfire, despite several other sites doing basically the same review, showing more than ample scaling on Crysis, basically they screwed up, purposefully or not, who knows.
Wide range of sites, I tend to compare/contrast between Anand, [H], bit-tech, another uk site that I believe still gets censored here, xbit sometimes, and a couple others, techpowerup.
But even then you have to be careful, I rarely ever read a conclusion or their opinion because you can find from the same results 10 different people will draw 10 different conclusions. Went through this in another thread, techpowerup who are maybe the most thorough site out there in terms of numbers/cards benched/games benched, round up the numbers to give overall performance at each res, where the 460gtx 1gb comes out only 10% behind the 5850.
When you look through the results the 5850 is very often 30 or 40% ahead, rarely less than 20% then a few games, a couple of TWIMTBP titles and a couple of cpu limited games(which narrows the difference because the games don't show the difference between say a 460 and 480gtx, even though we all know the difference is massive) and with those taken into account the end scores are much lower than in reality, they are.
Even then, whats fair, if the 5850 is 40% faster, but only 50% of games you ever play show it, and the other 50% don't show it, is it fair to say you'll only average a 20% difference, thats the problem, different reviewers will come to a different conclusion because their questions are different.
One reviewer is asking, how much faster is the 5850, another is asking how often does the 5850 show its extra speed.
Its very rare that reading 5-6 reviews(even bad ones) staying clear of the conclusions and reading the results(not any rounded up numbers or cherry picked results) you'll get a 95% accurate idea of how good/bad a card is.
Then theres other key issues, demo's vs full games, a review being honest and telling you what they did, how and why. Dirt 2 was a disaster on release as a LOT of sites only used the demo to benchmark, which wouldn't enable DX11 on the 480gtx, so we were getting AMD DX11 results vs Nvidia DX10 results. Maybe one or two sites did it on purpose but most genuinely just made a mistake.
When you're benchmarking its not always obvious a setting has taken effect, some games like Mafia 2 the new setting would only work after you exited the game and restarted it, can you tell if a card is using 8xaa, or 4xaa? Its a minefield to be honest.