I believe that the importance of data quality (ie: dropped, uncommitted, or orphaned records) depends on the application. That's why non-relational databases are more of a natural choice for some organizations than it is for others.

Let assume that Google, FaceBook, or Twitter had an intermittent transactional consistency issue such that 1 listings out of the top 100 were randomly excluded each time a user hit the website. Unless a QA engineer were systematically analyzing the results looking for such specific irregularities... no one would notice. Even if some users did notice, it wouldn't be a show stopper. It would be like "Oh, yeah we know about that bug and some people working on it.", but it's not as if external auditers or a regulatory agency is going to shut the business down until the problem is fixed.

In the banking industry, accounts have to balance, and information presented to clients is not subjective at all. Likewise, in the healthcare, government, or scientific industries, the data drives critical decisions and thus matters in a critical way.

For an e-commerce company like Amazon or e-Bay, those listing presented to users in the web browser are potential sales. I guess the same applies to Google and FaceBook, but to a lesser extent. However, from their perspective it's the paid add links, not the aggregated content, that matters most.