Speed claims, coverage maps and the proof you need

Suppose the network team sends marketing four numbers for the same fixed-line footprint: a peak observed download above 900 Mbps, a mean in the low hundreds, a median well under that, and a median measured between 8pm and 10pm that is lower again. All four are true descriptions of one network. Only one of them can go in the headline, and the choice is evidentiary rather than creative. Whichever number you print is the number whose measurement method you will have to reconstruct, with sample sizes and dates attached, long after the campaign stopped running.

That reconstruction is what this lesson is about. Who can stop the ad and on what grounds is settled elsewhere in the module. The question here is narrower and harder: what has to exist in the folder before the claim exists.

The number that goes in the headline

UK rules made the arithmetic explicit. For years, broadband "up to" claims cleared the bar if roughly one customer in ten could reachreachThe number of unique people exposed to your message in a given period. Unlike impressions, reach counts each person once, no matter how often they see it.View full definition → the advertised figure, which let carriers headline a speed nine buyers in ten would never see. From 23 May 2018, numerical speed claims in UK broadband ads have had to be based on the median download speed measured at peak time (8pm to 10pm), available to at least 50% of customers. Australia arrived at the same place from a different angle: the ACCC's broadband speed guidance pushed retailers to advertise typical busy-period speeds, measured between 7pm and 11pm, instead of the plan's theoretical maximum.

The shift from peak to median at busy hour has a consequence marketing teams often miss. Once the advertisable number is a congested-hour median, capacity spend becomes a marketing input. Adding backhaul that only relieves 3am load moves nothing you can print. Splitting an oversubscribed node cluster moves the headline.

A median can also improve for reasons that have nothing to do with engineering. If a batch of customers on slow legacy lines churns out, or a low-tier plan is withdrawn from sale, the surviving median rises while every remaining line performs exactly as before. Claim an improvement on that basis and the substantiation collapses the moment someone asks for the denominator. If you are advertising a gain, measure the same cohort across both windows, or state the population you changed.

Building a sample you can defend

Ookla's Speedtest data underwrites a large share of the world's network superlatives, and Ookla sells the product it measures with: Speedtest Intelligence subscriptions and Speedtest Awards whose badges are licensed for use in advertising. That does not disqualify the data. It does mean the vendor's methodology note is part of your evidence file, not background reading.

What a regulator or a rival's lawyer will interrogate:

• Who triggered the measurement. Consumer-initiated tests are self-selecting: people run them when something feels wrong, when they have just upgraded, or when they are sitting still on home Wi-Fi. That skew is not random and it is not constant between carriers.
• Device mix. A cellular median lifted by recent flagship handsets on mid-band 5G does not describe a base where a large minority still carry four-year-old phones.
• Time and day weighting. A sample that under-represents weekday evenings flatters every network, and flatters the most congested one most.
• Geography. Ookla and its peers weight by where tests happen, which is where people are dense. A national median can sit comfortably above the experience of every rural postcode in the footprint.
• Minimum n per region. "Fastest in the north west" resting on a few dozen tests is not a claim, it is a coincidence. Ask what the confidence interval is and whether your margin over the runner-up exceeds it. Overlapping intervals mean you tied.

Then there is the window mismatch, the most common failure mode in comparative speed advertising. An award covering the first half of the year is used in creative running in November, after a competitor has lit new spectrum. The claim was true when measured and false when broadcast. The ASA's position on comparatives is that the basis of the comparison has to be stated and verifiable, so the fix is mechanical: print the measurement period in the ad, and set an internal expiry date on the asset rather than letting it run until someone notices.

Coverage maps: the methodology is the claim

A coverage mapmapUsing software to automate repetitive marketing tasks and campaigns, enabling personalisation at scale across channels like email, web, and social.View full definition → is a model output presented as a fact. Two maps of the same network can differ by a fifth of the landmass depending on four choices: the propagation model and clutter data used, the assumed building penetration loss, whether a pixel counts as covered outdoors or indoors, and the throughput threshold at which "covered" begins. A map drawn at 0.2 Mbps outdoors is a different product from one drawn at 5 Mbps indoors, and only one of them matches what a customer thinks the colour means.

So the map has to disclose which map it is, and the legend has to be honest about confidence rather than painting predicted and drive-tested areas in the same shade. Publication date matters as much: US coverage reporting was criticised for years because a carrier could count an entire census block as served if it reached a single household in it, which is part of why the FCC moved to location-level data in its National Broadband Map (fcc.gov/BroadbandData).

The edge case that catches operators is technology retirement. When Telstra closed its 3G network in October 2024, the coverage story it had told for a decade was anchored in an 850 MHz 3G footprint. Higher-band 4G and 5G do not propagate the same way, so population and land-area figures inherited from the old network have to be re-derived rather than carried forward. Keep the two claim types separate while you do it: in Australia, a network reaching almost the entire population still leaves most of the continent dark, and a map that blurs population coverage into geographic coverage is the single easiest claim to challenge.

Also assume decay. Sites get decommissioned, spectrum gets refarmed, a mast loses its lease. A map published in January can be indefensible by June, which is why version history with dates belongs in the substantiation file next to the map itself.

What the substantiation file has to contain

Regulators do not ask you to prove a claim was true after the complaint lands. They ask what evidence you held on the day it launched. Who signs that file, and in what order across markets, is the gauntlet the sign-off lesson maps. The contents are this lesson's business:

1. Raw measurement data, not the summary deck: sample counts, dates, times of day, device and plan mix, and who ran the test.
2. The written methodology, including the definitions used for "coverage" and for the speed threshold, plus the propagation model behind any predicted map.
3. For comparatives, the third-party report as issued, its measurement period, and the stated margin of error.
4. Dated versions of every published map, retained for as long as the campaign is discoverable.
5. Reconciliation against what individual customers were sold. This is where Telstra was caught in 2017: the ACCC found NBN plans marketed at maximum speeds that the customer's own copper line could not physically deliver, and around 42,000 customers were offered remedies. The line-rate data existed inside the business. Nobody matched it against the plan sold.

Point 5 is the deeper lesson. Once a carrier can measure an individual line or cell, the proof obligation stops being a launch-day event. "Up to" is indefensible for a customer whose connection is already known to fall short, and the remediation bill for that is paid per account, not per campaign.

Choosing the claim your data can carry

The arbitration is not between bold and timid. It is between a superlative you can rebuild eighteen months later and one you cannot. A punchy claim with a thin sample costs the media spend already committed plus the withdrawal, and you cannot retrofit the evidence afterwards, because the network you measured no longer exists in that state.

The workable habit: pick the strongest claim your measurement design supports today, write the disclosure and the expiry date into the asset at the same time as the headline, and re-measure before the renewal rather than after the complaint. For the point-of-sale estimate a customer sees at their own address, the fair-disclosure lesson covers the duty; Ofcom's broadband speeds code of practice is the reference framework.

🎬 [VIDEO: "How Ofcom Regulates Broadband Speed Claims" - youtube.com/@Ofcom - search Ofcom's official channel for consumer guidance explainers on speed and coverage advertising rules]

Key Takeaways

• Median at busy hour, not peak: UK broadband speed claims since May 2018 must use the median peak-time (8pm to 10pm) speed achievable by at least 50% of customers, and the ACCC pushes typical 7pm to 11pm speeds in Australia.
• Interrogate the sample before the claim: who triggered the tests, which devices and hours are represented, how many measurements per region, and whether your margin over the rival exceeds the confidence interval.
• A rising median can come from churn or plan withdrawal rather than network improvement; hold the cohort constant if you are advertising a gain.
• A coverage map is a model, so the propagation assumptions, indoor versus outdoor, the speed threshold and the publication date are part of the claim. Technology retirements such as Telstra's 3G shutdown invalidate inherited footprint figures.
• Comparative claims need the measurement window printed and an expiry date set, because a true award used too late becomes a false claim.
• The substantiation file must reconcile marketed speeds against what the individual line can deliver: Telstra's 2017 ACCC case ended with roughly 42,000 customers offered remedies on data the company already held.