3DMark = Unreliable?

I see some members here that seem to discredit 3DMark as being unreliable, so I'm kinda asking specific reasons why some people discredit the software.

In my observation, 3DMark has it's purpose as a general performance tool. Yes, sometimes you cannot compare video cards between companies (Such as 8800GTX VS Radeon 2900XT, which seem to) , however, I'm quite sure you can use it to compare between very confusing video cards.

An example is say, GeForce 6500 VS 6600LE, most people can't say which is faster between these two. GeForce 8500GT VS 7300GT is another good one, since most hardware sites just review the fastest, no decent reviews happen in the low end sector, so for me, 3DMark is a lifesaver for buying the most performance on money.

Also, I'm also using 3DMark as an optimization software. Usually I'd try different settings and then test it's 3DMark score improvement, my idea is, if my 3DMark score improves, then general gaming performance improves, even if in reality, it could be just 1FPS difference. Is this true or not?

Last Q, kinda unrelated to my first question, but is there any explanation that the Radeon 2900XT and/or Radeon 3870 scores in 3DMark06 similarly to a 8800GTX, although in general gaming it is almost obvious the 8800GTX is faster?