Summary

  • Speed Score is a position on a publisher-held scale, not a rate: every metric is mapped to 0–1 by a logistic function "centering the values relative to a global reference" and only then weighted into a 0–100 figure.
  • An Ookla-produced Q3–Q4 2025 award report prints the component split as 70% download, 20% upload, 10% latency, while Ookla's own guide as republished by GSMA in 2020 attributed 90% to download and 10% to upload.
  • A market enters the Global Index when "the limits from the 95% confidence interval must be within ±5% of the median download speed" — the entry ticket is a precision test on the very quantity the table ranks.

A trophy on an operator's homepage is not a measurement. It is a claim used under licence, and behind it sits a figure called Speed Score. Three short passages of publisher text say what that figure is. None of them describes a speed.

The mapping

An average of megabits per second would at least stay in megabits per second. Speed Score does something else first. On Speedtest's awards methodology page the construction is stated in one sentence: "Speed Score uses a logistic function to adjust the values of each metric to a 0-1 scale, centering the values relative to a global reference." The same page describes the output as a 0–100 scale on which a score above 50 indicates performance better than the global average.

What that reference contains — which values, in what unit, over what population, for what period — the page does not say, and neither does anything else read for this article. A limit belongs beside the claim rather than at the end of it: a 0–100 Speed Score cannot be turned back into megabits per second from the documents read on 18 November 2026, because the one number needed for the conversion is the number that is not printed.

The blend

Inside each component, the central value is built from the percentiles of the sample the measured network itself generated. The awards page states that metrics are taken at the 10th percentile, the median and the 90th percentile and combined at a 1:8:1 ratio, so the median carries 80% of the weight. That page does not use the label "modified trimean". Ookla's own guide, as republished by GSMA in its 2020 revision, does, in one of the plainest sentences in the corpus: "Speed Score uses a modified trimean to combine speeds from the 10th percentile, 50th percentile (also known as the median) and 90th percentile in a weighted average using a 1:8:1 ratio, respectively."

The consequence is small arithmetic and a large idea. My arithmetic, A1: 0.1 + 0.8 + 0.1 = 1.00, so the combined figure sits between the two tails at a distance from the median of 0.1 × (P10 − P50) + 0.1 × (P90 − P50). A territory with a very slow tenth percentile gives up a tenth of that gap; a very fast ninetieth percentile recovers a tenth of its own. That describes the shape of the formula. It says nothing about any named network's numbers.

The weights, in two vintages

The component weighting is printed in two places in two forms, and the difference is not rounding. A Q3–Q4 2025 Belgium award report produced by Ookla and hosted by the winning operator carries the line "Speed Score Components: Download Speed: 70% Upload Speed: 20% Latency: 10%" (award report). The 2020 revision of Ookla's guide, republished by GSMA, attributes 90% of the final Speed Score to download speed and the remaining 10% to upload speed, on the stated reason that online experience is typically more affected by download. Two documents, two weightings, one name — so a current award claim and a chart drawn five years ago can describe two different instruments without either of them admitting it.

A second limit sits with this claim. The methodology guide on Ookla's own domain carries a last-updated stamp of 15 October 2025 and describes Speedtest measurement as consumer-initiated foreground testing, but the retrieved text ended before the aggregation and exclusion sections. The 70/20/10 line therefore rests on the awards page plus that verbatim composition line in an Ookla-produced award report, not on a full reading of the guide.

What counts once, and who is in the race

Two counting rules live on the same awards page, and both are about population rather than performance. Each sample represents the cumulative test results for each unique device or user per location per calendar day, so one handset cannot multiply itself by testing fifty times in an afternoon. And when fastest operators are analysed, only top carriers holding 3% or more of total test samples in the market for the entire award period are considered. The page does not print how many carriers clear that line in any market.

My arithmetic, A5: 3% is a share of that market's samples, not of its subscribers. In a market holding 10,000 samples the gate is 300 samples; in one holding 1,000 it is 30. The same operator can clear it in one territory and fail in another without changing a single radio.

The entry ticket

The Global Index does something the awards table does not, and it says so in one sentence. Country inclusion runs through a Precision Threshold, and on the Global Index page the rule is unambiguous: "Under this methodology, the limits from the 95% confidence interval must be within ±5% of the median download speed." Since June 2024 the reported values use a rolling quarter — the current month plus the two preceding months; since January 2025 mobile rankings use results from devices with Modern Chipsets; and fixed broadband includes mobile WiFi results.

The page does not print the estimator, the interval construction or the sample size behind that test. What it does print is a relative tolerance: ±5% of the market's own median. My arithmetic, A2: at a 50 Mbps median the permitted half-width is 2.5 Mbps, and at 300 Mbps it is 15 Mbps. Two markets face the same rule and not the same number. My arithmetic, A3: under a normal approximation, treating samples as independent, a half-width of 1.96 × σ/√n within 5% of the median requires n ≥ (39.2 × σ/median)², which is about 384 samples at a spread-to-median ratio of 0.5, about 865 at 0.75, about 1,537 at 1.0 and about 3,458 at 1.5.

My arithmetic, A4: because the tolerance is denominated in the median, the gate rewards a tight distribution and a large sample rather than a fast network, and a fast market and a slow market with the same relative dispersion need the same count. All of that is a scale for the gate. None of it reconstructs Ookla's own calculation.

Then a reading, offered as a reading: a market that fails this test is absent from the table not because it is slow but because its measured median was not tight enough. Absence and slowness are different facts, and the index's own admission rule is the reason they can look identical from outside.

The badge

The award is a designation; the mark that carries it is licensed. Ookla's 2024 brand licensing guide states: "Any usage of our brand marks or award assets must be reviewed for approval by Ookla." The same guide requires the trophy to appear in the same piece of collateral as the award claim or grouped with it, forbids overlaying, blending, recolouring, rotating or distorting the trophy graphic and forbids co-branding it with another logo, permits paraphrase of claim wording only in secondary references, and permits translation only on the basis of the wording Ookla provides in the Awards Certificate (brand guide).

So the public object — a line on a homepage, a trophy beside a price list — is the far end of a chain that begins with a logistic mapping against an unpublished reference and passes through an approval step. That describes a documented workflow. It alleges nothing about any operator that followed it.

What this reading does not say

Nothing here shows that a figure was falsified, that an award was purchased, or that a named network fails to deserve its claim. What the documents show is narrower and more useful. The score is a position against a reference held by the publisher. The components have been weighted differently in documents published five years apart. The global table's admission rule is a precision test on the same median it then ranks.

A limit belongs here too: no dedicated published award licence, fee schedule, term, territory or revocation clause was obtained in the documents read, and the brand guide is a visual design guide rather than the award contract. That is a statement about six documents read on 18 November 2026, not a statement about what exists. One requested address, speedtest.net/about/knowledge/methodology, returned HTTP 404 and is excluded; nothing here rests on it.

The trophy is real. What it certifies is a position, and the ruler that fixes the position is kept by the party that hands out the trophy.