Short version - with the Boost 1.0 methodology they limit the boost to the "guaranteed" level (as in the number quoted on the box) as with older Kepler cards they'd often fail to sustain their initial boost clock indefinitely - which for a short benchmark run would often show them massively ahead of the equivalent AMD card but in actual gaming fall down quite a bit - whereas Kepler 2.0 cards will usually sustain it all day. (Some places tested by looping the Kepler cards for an hour first before benchmarking - others would limit the boost clock).
i.e. my 780GHZ has a rated boost clock of 1072MHz but actually boosts to 1188MHz out the box (and never drops) which makes a fair difference to benchmark scores.
EDIT: It is actually more complicated than that though - as if they are using an older revision 1 780 that they've had sitting around since they originally tested it on release and if they have the boost clock set to the reference boost of 900MHz that is way down from what someone with a 2nd revision 780 would be seeing in the realworld where they are probably holding a 11xxMHz boost out the box.
Thank you