One thing I'd make explicit in the dataset docs: are the keyword/country rows requested independently, or as a shared comparison? Trends' 0–100 scale is relative to the query context. Two independently normalized series peaking at 100 don't imply equal demand, and the same applies to comparing countries. An optional shared anchor term, with the exact geography/timeframe/category/search property retained, would make the competitor-comparison use case much safer.
On the reliability side, distinguishing a legitimate low-volume/empty series from a transport failure would help consumers too. Otherwise a rerun can quietly turn “no data” into zero interest. I'd keep a status alongside the series and cache successful keyword/parameter combinations so a retry of a large batch only does the missing work.