Idle baseline
The browser sends 12 lightweight requests while the connection is not deliberately saturated. This establishes the unloaded median and distribution.
How LinkSpeed measures responsiveness while broadband is busy
This page documents what the live test measures, how the result is calculated, what the grade means and what the browser test cannot prove on its own.
Method summary
Browser request latency + controlled download and upload load
All large test payloads travel directly between the visitor's browser and Cloudflare's speed-test service. LinkSpeed serves the interface and stores only the small anonymous result summary used for aggregate benchmarks.
EndpointCloudflare speed-test service
Latency typeBrowser HTTP request RTT
Idle samples12
Load phases~7.5 s download + ~7.5 s upload
Workers4 download + 4 upload
Primary comparisonLoaded median − idle median
The four test phases
Download load
Four concurrent download workers generate traffic while request latency continues to be sampled. The phase lasts about 7.5 seconds or stops earlier if the safety cap is reached.
Upload load
Four concurrent upload workers generate upstream traffic while latency continues to be sampled for about 7.5 seconds or until the safety cap is reached.
Compare and grade
LinkSpeed calculates median, P25, P75, P95, range and jitter for each phase. The larger median increase from download or upload sets the grade.
How added latency is calculated
Loaded median − unloaded median = added latency
75 ms − 20 ms = +55 ms
If the quiet median is 20 ms and the download-loaded median is 75 ms, the download phase added approximately 55 ms of delay.
LinkSpeed uses the median rather than relying only on the mean because one unusually slow request can distort an average. The P95 and full distribution remain visible so occasional spikes are not hidden by a good median.
What each latency statistic means
Median
The typical middle response for the phase and the main value used for added-latency grading.
P25 / P75
The middle 50% of samples. A narrow range usually indicates steadier behaviour.
P95
A high-percentile value that exposes occasional lag spikes that a median can hide.
Jitter
The average absolute change between consecutive latency samples in the browser run.
Full sample range
Shows the best and worst observed requests and makes variability visible in the distribution chart.
Loaded latency grade thresholds
| Grade | Worst median increase | Interpretation |
|---|---|---|
| A+ | 0–5 ms | Exceptional responsiveness under load. |
| A | >5–25 ms | Good; only a small rise while busy. |
| B | >25–60 ms | Mild delay that may be noticeable in latency-sensitive use. |
| C | >60–100 ms | Noticeable queueing; verify over Ethernet. |
| D | >100–150 ms | High delay under load. |
| F | >150 ms | Severe delay under load. |
What this test measures — and what it does not
It measures
- Browser request round-trip time to Cloudflare.
- Latency while download traffic is active.
- Latency while upload traffic is active.
- Approximate browser throughput during the two load phases.
- Median, P95, jitter and sample distributions.
It does not prove
- That an ISP alone caused a poor result.
- ICMP ping to a game server, workplace or voice platform.
- Performance on every route or at every time of day.
- That replacing a router will automatically improve the result.
- That a Wi-Fi result represents the underlying broadband line.
A poor loaded-latency result is evidence of queueing somewhere between the browser and the Cloudflare endpoint. Repeat over Ethernet to determine whether Wi-Fi contributes before diagnosing router or ISP bufferbloat.
Why repeat tests can differ
Household traffic, browser scheduling, Wi-Fi conditions, VPNs, device load, provider congestion and internet routing can change between runs. Different testing services also use different endpoints, durations and traffic-generation methods.
- Repeat a poor result at least once with other downloads and uploads paused.
- Compare Wi-Fi with Ethernet.
- Repeat at a different time of day if evening congestion is suspected.
- Use the same device and connection type when comparing before/after router or SQM changes.
Anonymous result collection
To build UK broadband benchmarks, LinkSpeed stores the completed result summary: broad provider and city context where available, connection type, grade, median added latency, P95, jitter, throughput and sample counts. The stored result record does not include the visitor's IP address or browser user-agent string.
The large download and upload payloads are not proxied through the LinkSpeed web server.
Method ownership
Review and change control
Publisher: LinkSpeed. Method reviewed 15 August 2026. If sample counts, phase duration, grading thresholds or the measurement route change materially, this page should be updated alongside the live test.