Skip to content
Your files stay on your device

File processing runs in your browser. Documents are not uploaded to our servers. The site may download processing engines and send usage or feedback events without file contents.

How we handle your data

How our internet speed test works, and what it measures

Parallel downloads, growing uploads, latency under load: what the Laboratory speed test counts, and why the test server can change your result.

By LaboratorU

On 4 October 2026 we ran the same Chromium on the same 200 Mbit/s fibre line, again and again. When the nearest Cloudflare node answered, upload read 78 to 88 Mbit/s over HTTP/2. Whenever a node in a neighbouring country answered, it fell to 17 to 21, with either protocol. Nothing on the line had changed. The only difference was which Cloudflare node took the request.

That one observation shaped the page: the speed test shows the node city next to the numbers, and it warns when the node is in another country than the visitor. This article describes what the test does and where it stops being informative.

What the page runs

A full test has three phases, in this order: latency, download, upload. The code is in packages/network and the page logic in packages/ui/src/network/speed.

Latency. We send repeated small GET requests to /cdn-cgi/trace. The Cloudflare node answers that path itself, without a Worker and without a file lookup. Our first attempt used a static test file, and it gave 48 to 318 ms because of the asset storage lookup. The trace path gave 19 to 22 ms, which matched the TCP connect time. The result is the median of 20 probes (10 in the quick test), plus jitter, which we define as the mean absolute difference between consecutive probes.

Download. Three parallel streams read random, incompressible files of 1, 5 and 25 MiB served by the same site. We count the bytes that arrive, not what the browser reports, because WebKit reported zero sizes for these requests. The first 2 seconds of a 10-second run are thrown away so that TCP slow start does not drag the average down. The run also stops at 300 MiB.

Upload. The browser sends POST requests to the Worker in apps/net-worker. The Worker reads the body, counts the bytes and returns { bytes, ms }. It sets no CORS headers, so other sites cannot use it, and it stores and logs nothing. Anything above 70 MiB gets a 413. Chunks start small and double until one request takes about a second, up to 64 MiB. Streaming request bodies failed in WebKit and in Chromium over HTTP/1.1, so we send whole blobs. Fetch gives no progress events, so each request is counted into the measurement window by how much of it overlaps the window.

Latency under load. During each transfer a probe goes out every 200 ms, and only probes inside the measurement window count. The bufferbloat grade is the larger of the two increases, loaded download latency or loaded upload latency, minus idle latency. Under 5 ms is A+, under 30 A, under 60 B, under 200 C, under 400 D, and anything more is F.

What we compared it with

Date Setup What we saw
4 Oct Same line, node switching between the local one and one abroad Runs on the node abroad: upload 17 to 55 Mbit/s, latency 46 to 51 ms
4 Oct HTTP/2 only, local node Upload 78 to 88 against fast.com at 70 to 79
4 Oct Chromium, HTTP/3 on, local node Upload 52 against fast.com at 110
5 Oct M-Lab NDT7, three alternating pairs M-Lab download 77, 100 and 129 Mbit/s against 212, 303 and 308 on our page

The HTTP/3 row decided something concrete. On the same node, Chromium over HTTP/3 uploaded at 52 Mbit/s while the reference showed 110. Over HTTP/2 the two agreed. We switched HTTP/3 off for laboratoru.com on 4 October. Download was not affected by the protocol.

Our download numbers came out higher than fast.com in every pair (216 to 299 against 150 to 180). We do not know whether fast.com is limited by its own servers or we overestimate. A third reference was missing, so the page copy makes no claim about that gap.

Choices we made on the way

The M-Lab check is an optional second test under the full speed test. It uses a single TCP stream and ran 2 to 3 times lower than our page on download, which is what a single stream does by design. Our page, like most speed tests, opens several streams. The on-page note says so and gives no fixed percentage.

On slow lines the first version showed no upload value at all. The first upload request was at least 1 MiB, and three of them in parallel did not finish inside the quick window on uplinks up to about 3 Mbit/s. On 6 October we lowered the floor to 64 KiB and scaled the first chunk from the measured download speed. Regression tests now run 6 and 2 Mbit/s quick runs and 1.5 and 0.5 Mbit/s minimal runs. Each of those fails with the previous floor.

The Cloudflare rate limit shaped the upload client too. Our first rule allowed 20 requests per 10 seconds, and the test itself got 27 responses with status 429. We raised the rule to 40 per 10 seconds and reduced the request count with larger chunks. A 429 response is never counted as speed; it marks the upload as limited. On a local preview with a throttled line the most we saw in any 10 seconds was 19 upload requests.

What did not work

XHR with 5 MiB chunks gave 17 to 18 Mbit/s because it fires too few progress events, while fetch with growing chunks gave 46 to 51 against fast.com at 49, so we dropped XHR. The static file for latency read 48 to 318 ms where the trace path read 19 to 22. Neither Ookla nor speed.cloudflare.com would start from an automated browser, and we did not work around their bot checks, so the automated pairs used fast.com only.

One of four Chromium runs on the production preview showed “no data” for download, with no download requests at all. It did not repeat in three reruns. We still list it as unexplained.

What the numbers mean

The test measures how fast this browser moves data to and from the nearest Cloudflare node. It sees nothing of your router, your Wi‑Fi or other devices. Latency is an HTTP round trip, not an ICMP ping, and a game or call server farther away adds its own delay. Packet loss is not measured. If the node city is in another country, the upload can read far below what your provider sells you, as the runs on the node abroad showed.

The test moves up to 300 MiB in each direction in full mode. The page shows the data use before you start. The quick and minimal modes cap at 100 and 25 MiB per direction.

How we measured

Everything above comes from our own runs on 4 to 7 October 2026: one 200 Mbit/s fibre line, Chromium and WebKit, the production Worker, and a local preview server that can simulate bufferbloat. The test files are docs/domains/network-quality.md, packages/network/src/measure.ts and packages/ui/src/network/speed/runner.ts. All comparisons were back-to-back pairs, because the line itself varied during the day: fast.com showed 49 Mbit/s upload one morning and 74 to 130 later.

Try it

The Internet Speed Test is in the Laboratory. The download-only version uses less traffic and works even when the upload Worker is off.

Sources