A few guesses: 
 
To be considered observable, there are a few conditions that data should  
met, I believe: 
(1) Replicability, meaning that somebody else can reproduce your number  
without having to excercise value judgements. In that sense, CPI is an  
observable given that there is a set of instructions that tells you how to  
estimate it. 
(2) Availability, meaning that it should be relatively simple (not cheap,  
just simple) to collect 
(3) Consensual or relevant, meaning that relevant actors should agree about  
its relevance. For example, the CPI basket may not reflect individual  
baskets but in labour negotiations parties usually agree that it is the  
appropiate measure of purchasing power (a few arguments are mixed here:  
economics of information and "star" indexes) 
(4) Stability, meaning that you can "store" the record in a precise way. 
(5) Ordinality or cardinality, meaning that you have to be able to construct  
some kind of scale or state, even a binary one. 
 
More speculation: 
 
- The line between O and U is not fixed. My guess is that we will be able to  
measure "happinesss" at some point in a more "hard" way that through surveys  
(and "happiness" measured through surveys can only be considered data since  
a few years ago, I guess). 
 
- According to the above suggested conditions, it is not clear that you can  
reveal preferences neither through surveys nor through choice. For example,  
do I prefer to smoke or not? I might answer no but still smoke. Preferences  
then will be U no matter how much do you revealed by choice. 
 
I stop here. As with any other methodological question, is hard to answer  
without knowing what you are after. 
 
Best, 
 
Javier