The source is the Speedtest by Ookla Global Fixed and Mobile Network Performance dataset, published quarterly under a CC BY-NC-SA 4.0 licence. Ookla restricts this dataset to Speedtest results carrying GPS-quality location information, a subset of all tests taken, and aggregates them onto a fixed Web Mercator grid at zoom level 16 - tiles of roughly 610 metres at the equator, narrowing with latitude. Each published tile carries an average download speed, average upload speed and average latency, together with a count of tests and a count of devices, where devices are de-duplicated within a tile but not across tiles or quarters. Ookla's wider methodology also describes averaging each user's results into a single sample before aggregation, so that frequent testers do not dominate. Historical quarters are periodically re-aggregated, which is why the ingestion step tracks object versions rather than treating published files as immutable.
Quarterly tile files since 2019Q1 are synchronised from Ookla's public data storage (S3 bucket). Quarters Ookla re-aggregates after first publication are refreshed rather than assumed static, with change detection based on the object ETag. Each tile's quadkey is decoded to its centroid coordinates and assigned to a country by point-in-polygon test against the World Bank Official Boundaries Admin 0 layer (August 2026 vintage), identified by its WB_A3 code. Ties on shared borders are resolved deterministically, centroids falling just offshore were snapped to the nearest country within 1,000 metres measured in an equal-area projection (EPSG:6933), and tiles matching no country were excluded. The resulting quadkey-to-country assignment is cached and versioned by a content hash of the boundary source, so each tile is spatially joined once regardless of how many quarters it appears in.
Annual country level datasets are harmonized into the SDMX structure fitting the WBG's Data360 DSD. Country values are published as two statistics, distinguished by the composite breakdown dimension. The mean weights each tile's average download speed by its test count, sums across the country's tiles and divides by the country's total tests. The median orders the country's tiles by speed, accumulates test counts along that order, and takes the tile average at which the cumulative count reaches half the country's total. The two are not the same quantity measured twice. Weighting tile averages by test counts recovers the mean across the country's tests exactly, as those tests are represented in Ookla's published tile averages; the median, taken over tile averages rather than individual results, only approximates the test-level median and can fall on either side of it, because averaging within a tile compresses both tails of the underlying distribution. The mean is also the more volatile of the two at low sample sizes, where a handful of extreme tiles can move it appreciably. Speeds are converted from kilobits to megabits per second.
Annual figures pool all tiles from the year's quarters and apply each formula once, rather than averaging four quarterly results, so every quarter contributes in proportion to its test volume. A country-year publishes values only where it is supported by at least 20 speedtest results and the country appeared in at least two quarters of that year; the thresholds are applied to the country-year as a whole, so a row that fails either one publishes neither statistic. Observations failing either condition are withheld entirely rather than published with a quality flag. Values for the ongoing year are published where at least two quarters of data exist, carrying provisional observation status.