In the first session, participants saw two triples of restaurants, one that was used to devise a test pair for the second session and one that was used as a control pair. For simplicity, we label the three alternatives in each triple A, B, and C. For the first triple, participants chose between A and B. Then the loser of this pair—say, B—was compared with C. After choosing between B and C, participants performed a matching task, changing C (creating Restaurant C') so that it equaled B in overall attractiveness. Because B and C' were equally attractive and A was preferred to B, this process yielded a pair for the second session, A and C' , in which one alternative (C') was known to be inferior to the other (A).

So the second comparison wasn't between the original two choices. Instead, subjects were asked to modify a third choice to be "equal" to the one they rejected, and it was this third choice that got reworded and compared with their original preference. Restaurant preferences aren't all that solid to begin with, so I wouldn't read too much into this sleight-of-hand.