The phone screen would typically focus on a coding question that candidate would work on through a Google doc. The code was expected to be real code, not pseudocode, but it wasn't expected to compile. The questions weren't that hard either, e.g., determine if two strings were anagrams. More than a smoke test (i.e., not just FizzBuzz), but not scary either. The onsite interview was a full day in a single room, broken up by lunch. Interviewers would decide which questions they'd use, and they typically wouldn't coordinate in advance. Instead, they'd fill out a printed sheet of paper that stayed in the room with what general area they'd covered (e.g., Algorithms, Coding, Design) and which problem they'd used (e.g., Invert Binary Tree). Interviewers did not give any indication of the candidate's performance on the sheet. My favorite problem was string segmentation, which Google should now have on its list of banned questions (cf. Retiring a Great Interview Problem).