Drawing inferences from revealed preferences can work if individual preferences meet four seemingly simple assumptions (Hal Varian's Microeconomics Analysis provides the clearest exposition I know of micro theory. There is also a nice discussion here).
(1) Preferences are complete. This simply says that if I am faced with a choice of two consumption bundles, I can always say that one is at least as good as the other.
(2) Preferences are reflexive. This simply says that any bundle is always as good as itself.
(3) Preferences are transitive. This simply says that if bundle A is better than bundle B, and bundle B is better than bundle C, then bundle A is better than bundle C.
(4) Preferences are strongly monotonic. This simply says I never prefer less to more.
These may seem like mild assumptions, and they would be, except that people change their mind.
Perhaps there is a restaurant you go to on a regular basis. Its menu stays the same, and the prices stay the same over a reasonable length, so your choice set remains constant. Yet this week you might have the beef burger, next week the veggie-burger, and the following week the cajun chicken sandwich. In doing so, you have violated assumptions (1) and (3).
Except that, in a sense, you haven't. Suppose an element of the choice set is the time at which you eat your sandwich. Under these circumstances, the menu does change because it is offered at three different times. So we can preserve our theoretical assumptions.
But now if we want to estimate preferences with precision, we have a difficult problem, because we have to estimate in far more dimensions than the data can support. So anything we infer about preferences will necessarily be approximations.
This is not to say that the general axiom of revealed preferences isn't a powerful tool to learn things about the economy: Varian's lecture on the subject makes a pretty compelling case that it is. But it will always be an imprecise powerful tool.