Google search data ("Google Trends") is a powerful data source for cross-national and cross-temporal research into public interest and attention, especially in studying difficult-to-reach populations. But an often overlooked issue is that, as reported by Google, these data constitute sample draws from a vast population of queries. This article seeks to quantify the challenges posed by this sampling process, including the non-random nature of the resulting sampling variability, and identify the implications for global research. We collect dozens of Trends samples across a grid of 6800 distinct search queries with varied keywords, languages, time periods, and geographic parameters. We analyze the variation across individual draws for each query and find evidence that smaller population size and lower internet penetration are important predictors of sampling noise. Choice of language matters: even in settings where English is not the dominant language, searches for English language keywords frequently predominate. Finally, more historical time periods are noisier than more recent periods. These results demonstrate the importance of collecting multiple Trends samples, particularly for researchers studying regions with lower levels of development and Internet utilization. This study aims to highlight a data issue often ignored by social scientists and identify specific best practices to implement in future research using Google Trends.