It means that the RSA problem seems (at this time) to be more specific than factoring.
So the RSA problem is this: knowing a semiprime $pq$ and some exponent $e,$ and a value $v,$ find the $m$ such that $v \equiv m^e \mod pq$. (I actually got this wrong in my original answer, so that my phrasing of the RSA problem was equivalent to factoring up to some PP algorithm. Whoops! So you're not alone in being confused at the details here.)
The factoring problem is this: knowing a semiprime $pq,$ find both $p$ and $q$.
If you can efficiently solve the factoring problem, then you can efficiently solve the RSA problem: take the semiprime, factor it, use some theorems about prime moduluses to calculate an inverse exponent $d$ which reveals all ciphertexts as $m \equiv v^d$. (In fact these theorems are how the setup for RSA works: we know the two primes during the setup phase.)
However, it is not known that solving this above problem for arbitrary messages $m$ will tell you anything about the factors of the modulus or the exponents involved. It might or it might not; we don't know. Many smart people have presumably looked at the problem but nothing obvious has jumped out at any of them. So it's not known that the factoring problem is solved by solutions to the RSA problem (plus polynomial effort), only that the RSA problem is solved by solutions to the factoring problem (plus polynomial effort).
In fact in 1998 Boneh and Venkatesan published a proof that a certain simple class of algorithms (plus, times, exponents, no XOR/NAND type stuff) cannot be used to turn an RSA-problem solution into a factoring algorithm. The argument had a simple ingenuity to it: by manipulating those arithmetic operations mathematically, we can find out that the "reduction algorithm" (for precision: this is the algorithm which uses an RSA "oracle" for a semiprime to factor that semiprime) turns out to be a factoring algorithm in its own right, so that we can modify it to a variant which makes no calls to its oracle. So we have a trichotomy: either (a) there is no such reduction algorithm, or (b) the reduction algorithm does not have a nice arithmetic interpretation or (c) factoring is polynomial-time just like the reduction algorithm was.