- Facebook selects your friends' postings according to your closeness to their point of view. In doing so, they simply try to help you in selecting the most interesting stories. You can question what is the algorithmic definition of "interesting" but nothing is prevented from being discovered and nothing is deleted. Indeed, you can still access the full list of postings just by selecting the "Most recent" order view.
- Google can indeed provide different views of their results to different users, for the same query and at the same moment of time. This can be either a form of personalization (e.g. again trying to adapt to different users' tastes), or the fact that you are served by different machines which have different dataset, or the fact that you are in a limited deployed experiment where different algos are tested. Again no scientific evidence that anything is concealed from knowledge or exposure. The information is still there and searchable but different views are provided.
- All the other big kids on the block follow more or less a similar set of principles.
Still, the idea of exposing in your ranked results something "surprising" and "un-expected" is quite appealing and it's a part of intense machine learning research. In ML jargon, this is the "exploration" part where you get something new for learning something unknown and never observed before. Exploration is opposite to "exploitation". In Exploitation, you use what you have learned from the past to personalize your future flow of information.