Seven years of US internet speed tests, county by county: who caught up, what still explains the gap, and what the tests reveal about outages
Key results
Ookla publishes the results of Speedtest runs as quarterly averages for a global grid of roughly 600-meter tiles. This study takes every quarter from to for both home broadband (fixed) and mobile networks, keeps the tiles inside the United States, and assigns each one to its county: tile-quarters covering speed tests.
Three questions drive the work. How much faster did American internet get, and did rural counties share in the gains? What still explains why rural speeds are lower: the places themselves, their incomes, or how spread out their people are? And can a surge in speed tests be used to spot network outages?
Each line is the typical (median) speed test result in each quarter. Medians are used because a few very fast connections pull averages up sharply.
Counties are grouped by the USDA's 2023 Rural-Urban Continuum Codes: large metro areas (1 million people or more), smaller metro areas, nonmetro counties with a town of 2,500 or more, and rural counties with no such town.
Rural counties differ from cities in several ways at once: they are poorer on average, they sit in different states with different broadband programs, and their homes are far more spread out. To separate these, the analysis compares counties that are alike on one factor at a time and checks how much of the rural speed gap is left.
Speed tests can flag large or long outages. People run speed tests when their connection misbehaves, so an unusual burst of tests is a warning sign in its own right, and a stronger one when speeds drop or delays (latency) rise at the same time.
For counties with enough tests in every quarter, each quarter is compared with that county's own normal pattern. A quarter is flagged when tests run far above normal and either speed falls well below normal or delays rise well above it. That rule flagged of county-quarters. The strongest independent check is Hurricane Helene: both Buncombe County (Asheville) and Henderson County, North Carolina, are flagged in the fourth quarter of 2024, after the storm took down power and networks across the region in late September.
Ookla Open Data, performance tiles for fixed and mobile networks, Q1 2019 through Q2 2026, from the public AWS bucket (CC BY-NC-SA 4.0). Each tile is a zoom-16 Web Mercator square, about 610 meters on a side at the equator, with average download, upload and latency and the number of tests and devices. Tile centroids are decoded from the quadkey and assigned to counties by point-in-polygon against Census county boundaries. County attributes come from the USDA 2023 Rural-Urban Continuum Codes, the 2020 Census, and 2018–2022 ACS median household income, as compiled in the open-source rurality R package.
Speed tests are self-selected: people test more when something seems wrong, and the users who test are not a random sample of households. All statistics are weighted by test count, and medians are used throughout. Per-tile averages hide variation within a tile, so the 100/20 share describes places, not individual subscribers. Comparison windows pool four quarters to remove seasonality.
Figure 4 and Table A come from ordinary least squares regressions at the county level, restricted to counties with at least 200 tests in both 2019 and the latest window, with state-clustered standard errors. Figure 5's convergence slope comes from regressing log growth on log 2019 speed with income, rurality and state controls. The disruption detector removes the national quarterly mean from each log series, fits each county's own quadratic trend, and scales residuals by 1.4826 × the median absolute deviation.