They know your approximate location when you conduct an offline credit card transaction, and the approximate time and location of the store where the transaction occurred.
And while one such data point pair isn't conclusive, getting, say, a dozen matching pairs might identify a user with high enough correlation to be useful.
