I suspect this is a bit of an ambiguous term.

Obviously not all text is equally comprehensible to computers and humans alike. Perhaps the best definition of "computer readability" would be the ease with which a computer may infer the same sort of data that a human could infer by reading it (a caveat exists for things which are theoretically possible for a motivated, tenacious human, but too laborious for anyone to attempt - such as enumerating the exact counts of every letter in a book).

The emphasis on recapitulating human comprehension is relevant for two reasons:

1. For the vast majority of time in which knowledge has been generated (at least five thousand years for written language, even longer if you include unwritten knowledge), that knowledge has been generated exclusively by humans, exclusively for humans. "Reading a text like a human would" is vital to accessing that past knowledge.

2. Today, enormous amounts of knowledge are still generated by humans, for humans- in natural language (books, articles, drawings) that are not annotated or structured as a computer-compatible data format at all. The most important knowledge is still not stored in some computer-friendly database (imagine what a nightmare it is to answer a simple question like "in what direction does information flow between DNA, RNA, and proteins?" by computationally accessing published knowledge, when it takes a school boy 5 minutes of googling to come up with a very good answer), and probably will never be, because often experts of a field are not also experts of designing and using computational data structures.

In principle, one could also worry about humans understanding computer knowledge, but I don't think anyone's ever written a program that creates knowledge too complex and advanced for a reasonably educated human mind. Perhaps one day AI research will progress so far as to change this.

With the concept of "computer-friendliness" defined with some clarity, we can get to your question of what is "computer readable" and what is not. The thing is, nothing is perfectly computer friendly anyway. Even a simple 2D table of numbers saved as plaintext still requires some parsing to load into a program: The difference is always one of degree. All knowledge is computer readable, in principle, it's just a matter of how many minutes, days, years, or centuries it will take you to write an adequate parser. Conversely, you'll see it follows that no knowledge is computer unreadable (unless it's nonsense - I suspect that "knowledge" is defined as "human readable knowledge" in the first place). Anything a human can read, is probably (assuming AI research is not doomed) also computer readable given that sufficient effort is invested into reading it ("sufficient effort" may or may not imply extreme amounts of programming, which for instance may take longer than the lifespan of the universe, but that is tangential to the main point).

So, ultimately, there's no such thing (except perhaps if you subscribe to metaphysical philosophies of knowledge) as "computer un-readable knowledge", and "computer readable knowledge" is a tautology or redundancy, because all knowledge is computer readable and there is no other kind. There is only the question of how many programmer-hours you are willing to invest to programmatically access that knowledge.

Semantic Web is a good example for this concept.