Because "speed" is not a single, well-defined quantity, different tests can give materially different answers for the same connection. Cloudflare's Speed Test is deliberately built to describe the quality a connection realistically provides rather than a peak bandwidth figure.
What a speed test actually measures
What people usually mean by Internet "speed" is throughput: the rate at which data moves between sender and receiver, expressed in Mbps or Gbps—the units ISPs use to market packages such as 500 Mbps or 1 Gbps. Following that convention, most speed tests keep sending data until the connection's maximum throughput is reached.
Tests diverge in many other ways, and each choice changes the number a user sees:
- Where the test server sits in the network. A user in New York City testing against New Jersey and against Singapore will get noticeably different results even with identical methodology, because of distance and the links traversed.
- How the test decides to send more data.
- How many TCP/UDP streams it opens.
- Which congestion control algorithm it uses.
- How it aggregates the collected samples.
Latency, packet loss and sometimes jitter are commonly measured too, though these are less intuitive for end users than throughput.
Speed tests became popular when bandwidth was the main obstacle to a good experience. As bandwidth has grown—in some regions to the point of near-plenty—attention has shifted to latency under load (bufferbloat) and packet loss as the metrics that most degrade video calls and gaming. Speed tests nonetheless remain useful, not least because they are the best-known form of Internet measurement.
Inside Cloudflare's test
Results begin appearing as soon as the page loads: the browser immediately starts sending data requests to Cloudflare's Network Quality API and timing each exchange. The API runs on Cloudflare's global network on Workers, and anycast routing sends each user to the nearest data center.

Rather than saturating the connection, the test transmits a series of payloads of predefined sizes—data blocks—each a fixed number of times, then presents the results as box-and-whisker plots showing observed ranges and percentiles.
Each result combines the time to establish the connection with the time to complete the data transfer, minus any server "thinking time". Connection setup is governed by latency, since it requires packets to travel back and forth. Transfer time is bounded by bandwidth, congestion, server limits and the amount of data moved—and, perhaps surprisingly, small transfers are also throughput-limited by latency.
Latency, jitter and loss under load
While throughput measurements run, the test fires empty requests at regular intervals to capture loaded latency: the round-trip time to Cloudflare's network and back while the connection is busy. This contrasts with idle latency, recorded first as the page loads, which represents the lowest latency to expect. The test also reports idle and loaded jitter—the average variation between consecutive RTT measurements, indicating stability—and packet loss, the share of packets that fail to arrive when relayed through a WebRTC TURN server over a period of time.
Aggregate figures for each metric update live during the run, but nothing is finalized until every sequence completes. The full measurement set then feeds the Aggregated Internet Measurement (AIM) score, which recasts performance in end-user terms such as suitability for streaming, gaming or video conferencing. The AIM score is a summary; the detailed results are what reveal—and conceal—about a connection.
Interpreting the numbers
Cloudflare's test runs on Cloudflare's own global network. Other providers put servers close to end users or major exchange points to capture performance under specific conditions; Cloudflare measures in the context users actually encounter daily—fetching content served through its network. Any test built on the Network Quality API behaves the same way.
Because the test does not try to saturate the connection, its download and upload figures are not maximum throughput but the rate at which you can reliably expect to move various data sizes. The distinction is small but meaningful: the aim is typical delivered quality, not peak capability.
That framing matters because everyday traffic rarely maximizes bandwidth. Video conferencing, streaming, browsing and even gaming need little bandwidth and are far more sensitive to latency, jitter and loss. A high score on a saturating test—mirroring the service level purchased from an ISP—does not guarantee a high-quality experience. Which metrics dominate depends on the use case: a gamer gains more from lower latency, a remote worker from lower jitter. For most modern workloads, throughput is one contributor among several.
The test also runs primarily in the browser on the end-user device, so its measurements fold in bottlenecks beyond the access network—the browser itself, the local Wi-Fi network and similar factors. Results reflect the combined performance of every component along the path to content, not the ISP alone.
Users often run speed tests to check whether they receive the service they pay for. That is a reasonable question, but no standardized definition exists for answering it, so no test—Cloudflare's included—is a definitive measure of ISP service. It is a useful way to assess quality of experience when accessing content delivered by Cloudflare's network.
Reading the results: two contrasting connections
A test that scores “Great” on all three network quality rubrics points to consistently good quality, while a “Average” score suggests reasonable quality with occasional glitches. Scores can also diverge across video streaming, online gaming and video chatting, depending on how balanced the connection is for those cases. To explain the scores, the numbers have to be examined.
A residential fiber line accessed over Wi-Fi

This run used a laptop on Wi-Fi in a single-family home with a 500 Mbps fiber connection. Download topped out at 406 Mbps, short of the contracted rate. Wi-Fi is usually the bottleneck on high-speed links and a frequent source of instability.
The relatively consistent shape of the download and upload graphs, together with low jitter, suggests good reception and little activity from neighboring Wi-Fi networks — the two most common causes of poor Wi-Fi. Latency sits well within what is expected in an urban area, and a wired connection could shave about 2 milliseconds off it. The gap between idle and loaded latency is small, so a video call should survive background file synchronization without glitches. Large idle-to-loaded differences are a common warning sign: if they approach 100 milliseconds or more over a wired connection, the ISP is likely at fault.
Higher-bandwidth connections should display lower idle to loaded latency differences. The higher the bandwidth, the less likely it is to be fully utilized in practice. However, congestion further upstream in the network can drive these numbers up, especially if your ISP is oversubscribing its capacity.
The slow start visible on both graphs comes from the way transfers ramp up: senders progressively increase how many packets are in flight per acknowledgment, beginning with one. Small transfers are therefore limited by latency, since each acknowledgment round trip delays the next data packet. This is visible in the per-size charts — transfers smaller than 10 MB cannot use the full bandwidth of this connection.

That is one reason pure speed is no longer the main indicator of quality of experience on modern broadband: ordinary browsing consists largely of small transfers that cannot saturate a fast link.
The same laptop on cellular 5G

Here the picture inverts: speeds are much lower and inconsistent over time, latency is higher — particularly under load — and latency jitter is quite high. The modest speeds hint at a location that is not densely populated; dense 5G coverage should bring higher speeds and lower latencies, though more simultaneous users in such areas drive speeds down and latencies up through congestion. The detailed latency charts show irregular values, some above 100 milliseconds.

Quality and convenience pull against each other: reaching the Internet from anywhere in the house, a park or the beach carries tradeoffs. The reports let users understand those tradeoffs, compare against peers or other providers, and choose with more information.
Why Cloudflare runs a speed test at all
The test aims to give end users insight into their connectivity and to improve the Internet by making its performance transparent. Its engine is open source, so anyone can run it for their own research and verify how results are produced. Every result is published to Measurement Lab's public Internet measurement dataset in BigQuery and is also reachable through Cloudflare's Radar API, giving researchers, policymakers and network operators data for analysis. Results are anonymized to protect user privacy and are never used for commercial purposes.
What's planned next
The test dates from 2020 and has become a go-to resource for measuring end-user network quality, with much positive feedback on its understandable interface and the metrics it reports alongside throughput. Planned work falls into three areas.
- Broader measurement. Cloudflare is expanding the reach and scalability of its Network Quality API so third parties can integrate it more easily and measure their users' connectivity over Cloudflare's network. UNICEF already uses the Speed Test in its Giga project to connect every school in the world, and Orb uses the API in its diagnostic measurement suite for continuous monitoring from any platform or device. Through 2026, Cloudflare plans to significantly increase the number of third parties powering their own tools with the Speed Test and API.
- New capabilities. More detailed performance analysis is in development, including support for higher throughput measurements — not the sole indicator of connection quality, but still important for diagnosing enterprise or shared-office networks where many users share one link.
- Better diagnostics. Users reach for speed tests both to verify the service they pay for and to troubleshoot. The goal is richer metrics and more advanced functionality to answer: where is the bottleneck? Is it in the local network or the ISP's? Does the issue affect only specific applications? Is it unique to one user or shared across a region?
Running a test
A Cloudflare Speed Test can be run at speed.cloudflare.com.



