Sustained measurements. Clear limitations.
Continuous transfer testing, automatic endpoint selection and latency under load.
Updated November 7, 2026
The measurement engine
InternetSpeed Pulse uses LibreSpeed’s open-source browser worker, with telemetry disabled and source/license notices available below the test. It is an independent service, not an Ookla product. Our integration adds validation of endpoint responses and rejects failed HTTP transfers.
Server selection
Choose Cloudflare’s measurement endpoint, our hosting edge, or Automatic. Automatic warms each available endpoint, takes three idle probes and chooses the lowest median response time. This compares only the available endpoints; it does not find every nearby server or force a specific city. Network paths still differ from an ISP-hosted Ookla server.
Continuous download and upload
Each direction measures a continuous 12-second transfer window. A 1.5-second download warm-up and a 3-second upload warm-up allow the connection to settle before counters reset. Single uses one request stream; Multi uses four staggered parallel streams. Completed requests are replaced while the window remains open. HTTP/2 or HTTP/3 can reuse underlying transport connections.
Preflight and byte accounting
Before the sustained phases, a small download verifies the expected binary payload and estimates a download request size between 1 and 32 MB. Uploads use reusable 4 MiB random-data blobs. The worker sums progress-event byte increments over elapsed sampling time, including network scheduling gaps. The overhead compensation factor is explicitly 1.0: no percentage is added to make the result resemble another service.
Idle and loaded latency
The engine makes twelve small idle HTTP requests. It reports the lowest eligible response time, using Resource Timing when available, with elapsed request timing as a fallback. Jitter follows LibreSpeed’s weighted variation calculation, which gives spikes more weight. Separately, lightweight probes during download and upload provide median loaded HTTP latency. Missing loaded samples remain unavailable. These values are not ICMP ping or latency to a game server.
Duration and data use
A typical run takes about 35 seconds, with a 75-second overall timeout. Data use depends on connection speed and can reach hundreds of MB or several GB; the previous approximately 37 MB quick-test cap no longer applies. Data charges may apply. Stop cancels active requests. Failed and stopped results are not saved or exported.
Comparisons and history
For repeatable comparisons, use the same device, wired connection, mode, endpoint and test conditions. The result now records the measurement engine version. Older fixed-payload results are labeled Legacy quick test in history and should not be compared directly with sustained tests. A stronger measurement method does not guarantee identical Ookla numbers or independent accuracy certification.
Connection detection
The speed test panel automatically reads the public IP and network organization from hosting request metadata. The IP is hidden initially, and IP/provider details are not stored in test history. Registration labels can describe a VPN or corporate network instead of your billing provider. Geographical and commercial labels are approximate.
Provider references
Cloudflare privacy policy
Cloudflare measurement endpoint documentation
Google Ads Settings