We build and maintain data tools on Apify: Google Trends, Google News, App Store and Google Play reviews, and bulk PageSpeed audits. Here we write practical Python how-tos.
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.
Thanks so much, really thoughtful points! 🙏 You’re right, by default each keyword/country is normalised on its own, so two peaks of 100 don’t mean equal demand. compareKeywords puts up to 5 keywords on one shared scale and keeps the geo, timeframe, category and property on every row. Beyond 5 keywords, and across countries, there’s still no common scale, so a shared anchor term is a great idea, and I’ll make that much clearer in the docs. Same with empty vs failed, failures already come back as separate, free error rows, but an empty series currently looks like 0 interest. A per-row status and only retrying the missing combinations are now on my list. Thanks again!
Aldo
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.