It is the size of memory area mapped to the AGP card.
It still seems to be inconclusive if it **INCLUDES** the memory on the AGP card or not - reading between the lines of some tests, I would say it DOES - in the resources for the card, there is only ONE large area of map.
Taking it on that basis - the amaount by which it exceeds the card's RAM, is the amount of System memory which MAY be used for AGP texturing if the card's RAM cannot hold all the textures in use.
Advice varies - Half system RAM, twice the card RAM (or the lower of the two). For some cards, 256 Mb is reportedly the "sweet spot")
http://www.technologyvault.co.uk/geforce/faq.php?display=faq&nr=25&catnr=2&prog=gef&lang=en
I find that for my Geforce 2MX anything higher than 128 appears not to be used - and at 128, I get faster performance during major scene changes in the test (when lots of new textures are being transferred).
This is a 32Mb card, and I have 384Mb RAM - so 256 breaks the "half system RAM" rule and MAY be limited - my old Rage128 though, limited to 64Mb - above that, the size of the memory region in the card's resources did not increase - and I take that as being the effective limit, and then lower the setting to that point.
There is a point towards the end of the Car chase in 3dMark 2001 - where you go under the second baddie as it falls and everyhing is lit up by it - THAT is where AGP aperture makes a difference - perhaps more so in high detail.
PS. My theory - give it as much as you can afford (the half system RAM breakpoint), no more than it can use (check the card resources, and calculate the size of the largest area), and it will use as much, or as little as it needs - unlike memory for onboard graphics, AGP aperture is NOT deducted from usable memory unless it's actually being used!