What I'm suggesting is that instead of CSV dumps websites should allow downloading micromodels -- user-related subsets of a bigger data model used by the web-service itself. Such micromodels would contain several linked normalized tables with only data relevant to the user who requested it. From a technical standpoint it can be one SQLite file generated on the fly. The SQLite file format allows packing multiple tables into one file which can hold millions of records.
Having a relational micromodel would allow more meaningful and interesting data analysis. It would play well with popular data analysis tools (except Excel which is poorly suited to work with relational data by design). Support for SQL queries would immediately make it compatible with vast amount of systems.
For information providers, an SQLite file with micromodel would be of size similar to current CSV dumps so it won't increase workload and traffic. Also generating micromodels can be even faster than generating CSV dumps since it won't require joining multiple tables in order obtain a denormalized view.
Below is an example of a possible micromodel schema for PayPal.