Short answer: What the documentation is saying is that if you need to keep an object that has been autoreleased in an autorelease pool, you need to retain it.
Long answer: For instance, say I need to do a certain operation to 1000 objects. Once I'm done with these objects I'm going to autorelease them. Without an autorelease pool, they're going to be eventually released, but holding those 1000 objects in memory can make your program really slow (at least until they're autoreleased).
In order to solve this issue, I'm creating an autorelease pool that will be cleaned every 100 objects. However, what happens if I need to keep the last object of the last batch around? I still need to purge those other 99 objects. What I'm going to do is send a retain message to that very last object then clean the autorelease pool.
This way the autorelease pool will notify the system that it no longer wants those 100 items, but you've already let the system know that you do need one of them. If the object had a previous retain count of 1, then it'll still be around:
1 (original retain count) +1 (your retain) -1 (autorelease pool release) = 1.
This preserves the object after the autorelease pool is done with it.