Background: The definition of “Infodemiology”[1] has grown to include using publicly available information (typically internet data), for health surveillance[2], with a dedicated infodemiology journal. A review found Google Trends (GT) and Twitter (now X) to be the most commonly used data sources[3]. A common misconception surrounding GT data is that it is “big data” because Google fields several billion searches daily. However, few researchers recognise that there are two distinct sources of GT data (Web vs API), nor reckon with the GT documentation which explains that results are based on a 10% sample of Google data.

Methods: Google permits one newly sampled extraction daily per search query. We gathered 2000 unique samples for daily and monthly data, and 500 samples for yearly data, for worldwide, Australian and US searches for ten different terms, amassing a dataset of 117,982,980 data points. We examined the effect of term probability, sampling variation, time aggregation (daily/monthly/yearly), region size, and data source on the variability of possible study results. We sampled mentions on Twitter (full data) for the same terms to provide a correlation measure to GT data.

Results: We found that aggregating multiple samples reduced variability across all other conditions—the major recommendation from this work. Terms with higher search probability showed less variability. Early GT data showed more variability than later data, presumably because of Google’s growth in search market share over time. More temporally aggregated data (yearly vs monthly vs daily) showed less variation. Searches from geographically smaller regions showed more variation. Correlations between Twitter data and individual samples varied widely, showing a significant impact on study results.

Conclusion: The implications of the findings extend beyond only infodemiological data, to, for example, other big data such as the Australian PBS 10% sample, used in many pharmacoepidemiological studies.