The automatic analysis finds errors in the push/fold stage of the tournament or in the hands which happen to be played as if they are in the push/fold stage already. Usually, such hands will involve a lot of folds and some short stacked players.
The ideal hands in terms of automatic analysis would be the hands from the late stage where all players are limited to just push/fold decisions. The threshold which automatic analysis uses is that the effective stack in the appropriate situations is less than 20 big blinds.
The effective stack is calculated as the second biggest stack between players who didn't open fold. Open folding would be folding before any player made a VPIP action.
So for example, if we are 9 handed and there are 8 big stacked players at the table and you're a short stack at the SB, and it is folded to you, then the effective stack is your stack (second biggest stack between you and the player at the BB) and such hand gets auto analyzed.
On the other hand, if the hero is short-stacked but there are several big stacked players behind him who are not limited to just push/fold and are very likely to play postflop against each such hand is not analyzed automatically, because not all players are in the push/fold stage yet.
Another potential reason for hand to get filtered out of autoanalysis is if you didn't push or fold in the hand.
If you raised non-all-in or limped, ICMIZER won't be marking it as an error or not, because these plays usually indicate a nonbinary strategy and stage of the tournament and applying simple binary analysis in such a stage would not be optimal.
The goal of autoanalysis is to be helpful and to minimize the chance of misleading the player. That is why the matching criteria for appropriate hands are pretty strict.