How to Read Broadband Speed Reports
A broadband report is only as useful as the population, measurements and calculations behind it. Before acting on a headline about national speed or a provider's rank, ask whether the evidence describes your plan, location and question. This guide is a reading checklist, not a new survey of internet speeds.
Ask what was counted
A count of tests is not automatically a count of people or households. One connection may contribute repeated runs; some locations or plans may contribute many more measurements than others. A report should explain what one observation represents, how repeat submissions are handled, the collection period and the population its conclusions cover.
The FCC's historical Measuring Broadband America methodology describes recruitment, test equipment, service-tier checks and data processing. The page now explicitly labels that material historical and notes the end of the former fixed-data collection partnership. It is useful as an example of methodological disclosure, not a source of current provider performance.
When those details are missing, do not assume a very large headline count makes the sample representative. Nor does missing documentation prove the underlying observations never existed. It means the claimed inference cannot yet be checked.
Check whether the groups are comparable
Provider averages can mix service tiers, locations, devices and connection methods. A comparison needs to explain which differences it controls and which remain. The historical FCC methodology documents tier validation and says a provider may be absent from a speed band when a representative sample could not be formed.
- Tier: are similar offers being compared, or does one group contain faster plans?
- Path: is the measurement wired, Wi-Fi, mobile or a mix, and what endpoint is involved?
- Time: do collection windows and busy/quiet periods align?
- Selection: who chose to test, which tests failed, and what was excluded?
These questions are reasons to qualify a conclusion, not to “correct” a result by silently discarding inconvenient observations. Keep exclusions explicit and preserve the underlying audit trail.
Read the statistic before the headline
A mean and a median answer different questions. A mean adds the values and divides by their number; a median is the middle of the ordered values, with the middle pair averaged for an even-sized set. A percentile describes a position in that distribution, but its finite-sample calculation also needs a stated convention. Do not call one of these another.
Check units and direction. Throughput, round-trip delay and variation in delay are different measures. A high download result does not establish that latency stayed low during a simultaneous upload. The FCC's historical test descriptions distinguish throughput, latency, packet loss and latency under load rather than collapsing them into one speed figure.
If a report offers a composite grade, ask for the inputs, weights, normalisation and missing-data treatment. A letter stored in a file is not evidence of the calculation that produced it. Grades from different tools are not interchangeable without matching definitions.
Keep advertised plans separate from measured experience
A plan document describes an offer; a test describes an observation under particular conditions. In the US, the FCC's consumer-label guidance helps locate plan information. It does not turn a household browser test into a national provider comparison or a legal finding.
For your own troubleshooting, Ofcom recommends repeated tests at different times and a wired comparison. Preserve the setup and failures as well as successful results. That produces a more actionable support record than comparing an isolated result with an unrelated national headline.
A short evidence checklist
- Origin: is the data collector named and is the measurement method available?
- Scope: are dates, population, geography and service tiers clear?
- Processing: are validation, duplicate handling and exclusions documented?
- Calculation: can someone reproduce the displayed count, statistic or score?
- Presentation: do the title, chart, caption and structured description say the same thing?
- Limits: does the author explain what the result cannot establish?
A failure on one item is a question to resolve, not permission to invent a substitute dataset. Privacy-sensitive raw records need not be made public: an appropriately scoped aggregate, method and auditable calculation can support review without exposing individual subscribers.
Use evidence for the decision in front of you
For a purchase, compare the actual plans available at your address using the ISP comparison checklist. For a slow connection, use the controlled measurement and support checklist. This page no longer publishes the earlier original-research findings or graphics; it should not be cited as a measured report on global or US speeds.