We have BigTable and protobufs and real MapReduce, so there's no reason to use MongoDB with its JSON storage and Javascript-based mapReduce() function. It's like comparing CSV to SQLite.
A major need for a company like Google is visibility into their distributed architecture. When you make every component be built on their libraries running on their architecture, you have a common toolset across all of your properties to track any performance problem back to its root cause.
MongoDB does not provide this type of visibility. Either within that common toolset or in any other way.