Download speed test

#1 What a download speed test actually measures

A download result is “data from the internet to your device, per second” — but that sentence hides a fork that explains most tool-to-tool arguments: how many streams the test opens.

Measurement Lab’s documentation of its NDT (Network Diagnostic Tool) is explicit about its design: NDT is “a single stream performance measurement of a connection’s capacity for ‘bulk transport’ (as defined in IETF’s RFC 3148),” reporting upload, download and latency (S53). One TCP connection, pushed hard, measured carefully.

Many consumer tools instead open multiple parallel connections and aggregate them. Neither approach is wrong; they answer different questions:

So when one tool says 300 Mbps and another says 180 Mbps on the same line, the tools are not contradicting each other — they are measuring a line’s ceiling and a connection’s capacity. The number that matters is the one that matches your question: “can one device stream 4K” is a single-stream question; “is my 500 Mbps plan delivered” is closer to a ceiling question.

#2 The FCC benchmark table

The FCC’s Broadband Speed Guide publishes per-activity minimum download speeds, based on running one activity at a time (S55):

Read the FCC’s own caveat with the table: “These numbers are rough guidelines and are not based on surveys or experiments conducted by the FCC. You should use your best judgment” (S55). They are ordering tools — 4K needs roughly 25, multiplayer gaming needs 4, a household doing both at once needs the sum plus headroom — not certification standards.

Google publishes a similar ladder for its own test, and the two agree at the edges: 100 Mbps and up for a whole household doing everything at once, down to under 3 Mbps where “web browsing should be fine, but videos may load slowly” (S51).

#3 How to isolate the download number

The method is the same one every provider documents on its own pages, assembled here once:

1. Go wired where the result matters. Verizon: connect directly to the router by Ethernet (S45); its slow-speed guide adds disconnecting VPNs to the checklist (S48). Spectrum’s speed-test article says it plainly: “For best results, use a wired connection instead of WiFi” (S38).
2. One device, everything else idle. Verizon says to turn off other connected devices during the test (S45); AT&T’s tip is to disconnect devices not in use (S49).
3. Close background traffic. Verizon names OS updates, antivirus updates and file-sharing apps as significant bandwidth consumers (S46); Spectrum names games, browsers and in-progress downloads (S38).
4. Same tool, same server, several runs. Tool and server choice change the measurement — the accuracy guide covers the mechanics; here it is enough to fix them and repeat, ideally at two times of day.
5. Log download, upload and latency per run. The pattern across runs is the finding; any single run is a sample.

One data-cost note worth knowing before you run tests casually: Google’s Search speed-test widget states that a test “usually transfers less than 40 MB of data, but may transfer more data on fast connections,” and that your IP address is shared with Measurement Lab when you run it (S54). On a metered or capped connection, repeated testing is real usage.

#4 Why your download result is lower than your plan

Providers print this expectation on their own pages — it is the first thing to check before diagnosing anything:

• Verizon: “factors like wifi strength and device capabilities mean that results are often lower than plan speed” (S45).
• Spectrum: results “may vary based on device capabilities, number of connected devices and router placement,” with a printed 4.7 Gbps maximum for any wireless connected device and “actual speeds vary and are not guaranteed” (S38).
• Xfinity: the app’s speed-to-gateway test exists precisely because device-side results include WiFi conditions and hardware limits; gaps are “most often caused by WiFi connections, rather than by the network itself” (S51’s provider-equivalent is documented on Xfinity’s own pages).

The compounding factors, in the order they usually bite: WiFi path (distance, walls, interference), the device’s own radio and age, other devices sharing the connection, the server’s path and load, and time-of-day congestion — Google’s guidance for its own test notes networks are “most crowded during peak hours like dinnertime” (S51).

A result that is below the plan but comfortably above every activity you run is, by the numbers, not a fault. The FCC table in #2 is the arbiter: 140 Mbps on a 300 Mbps plan still streams 4K five times over.

#5 Reading the number: what each range supports

Google’s activity ladder, published for its own test results (S51), is the plainest real-world scale:

Cross-checked against the FCC table (S55), the two sources agree within their tolerance: HD streaming lands at 5–8, 4K at 25, and “a whole active household” starts around 50–100. If your wired download number clears every row your household actually runs — with a little headroom for the row you might add — the download side of the connection is adequate, whatever the plan tier says.

#6 When a low download number is a fault

Work the fork in this order:

1. Wired or WiFi? If wired runs meet the plan and WiFi runs do not, the fault is in-home wireless: placement, interference, device age. No plan change and no provider call fixes that; the accuracy guide covers the split-testing pattern.
2. All hours, or peak only? Time-of-day patterns point at congestion (S51), not the line. Timestamp your runs.
3. Provider-documented threshold. Some providers publish one — Xfinity treats gateway results at 90% of plan as the healthy line (S33, cited on the Xfinity speed test page); Spectrum documents a wired-reset-rerun routine instead (S38). Use your provider’s number where it exists; the provider guides collect them.
4. Escalate with the log, not a screenshot of one run. A dated series — tool, server, wired/WiFi, times — is what turns “it feels slow” into a support case that goes somewhere. And stop before the factory reset: no provider’s download-speed documentation calls for one.

#7 What If Scenarios

Q: Why is my download much higher than my upload? Connection architecture. Cable networks are built download-first — Spectrum describes its network as “powered by fiber and connected to the premises by coaxial lines,” with the caveats that measurements fluctuate and speeds are not guaranteed (S38). Newer fiber products advertise symmetric speeds — gFiber’s plans, for instance, list “up to 1 gigabit per second upload (send) and download (receive)” (S52). The gap you see is the architecture, not a fault; the fault threshold is upload below what your activities need.

Q: Two tools give me different download numbers on the same connection. Which is right? Both, for their own question. A single-stream tool measures one connection’s bulk-transport capacity (S53); multi-connection tools measure an aggregated ceiling. Fix one tool and server for tracking, and use the other only for cross-checks — never average them.

Q: Does a speed test use much of my data cap? Google’s widget documents under ~40 MB per test, more on fast connections (S54). A handful of runs is negligible on unlimited plans — and Xfinity’s current plans include unlimited data — but on metered or capped connections, treat repeated testing as usage and check your plan’s terms.

Q: Is “download speed” the same as “internet speed”? Colloquially yes — Google notes download Mbps is what is “generally understood to be ‘internet speed’” (S51) — but upload, latency and jitter are separate measurements with separate failure modes. A line can pass every download test and still break video calls on jitter, which is why the AT&T guide treats those numbers as first-class results.

#8 Related articles

• Internet speed test accuracy — the full method: layers, wiring, server choice, patterns.
• Google speed tests — the Search widget, gFiber and the Google Home app test, including M-Lab’s single-stream design.
• Xfinity speed test — the 90%-of-plan gateway threshold and its documented exceptions.
• Spectrum speed test — Spectrum’s own wired-reset-rerun routine for low results.