The practical guide

How measurements are produced and compared

Published by InternetSpeedGeo · Reviewed

The question a browser test can answer

×

A browser speed test observes the path between one browser and a measurement service at a particular time. That path includes the device, its local connection, the router or mobile access network, the internet provider, intermediate routing and the measurement endpoint. The resulting number is a property of that complete observation. It is useful for troubleshooting, but it cannot automatically identify which component limited the result or describe every destination you use.

This distinction determines how to design a comparison. If you want to evaluate a room's wireless coverage, hold the device and test service constant while changing location. If you want to investigate evening slowdown, preserve the connection and device while changing the time. If you want to compare a plan with observed performance, first check that the local connection can support the plan's advertised rate. A result without these conditions can still describe what happened, but it supports fewer explanations. The methodology therefore combines implementation details with advice about comparison. The purpose is to make observations interpretable, not to turn an ordinary browser into a certified laboratory instrument or a contractual test of an internet provider.

Current measurement engine and destination

×

New full speed tests use the Cloudflare Speed Test JavaScript library. Download and upload requests go to Cloudflare measurement endpoints through the connection available to the browser. The endpoint is separate from the InternetSpeedGeo page hosting environment and from the destinations used by your games, calls or file storage services. The library supplies measured summaries, which the interface formats for display. The website does not fill an unavailable speed result with a representative provider value or a demonstration number.

A measurement service can be reachable through a path that performs differently from another service. Geographic distance, peering, congestion, connection reuse and service behavior can all contribute to differences. Consequently, a good Cloudflare result does not prove that a specific remote application is healthy, and a poor application experience does not by itself show that the speed test is wrong. Use the result as one observation with a named destination and method. The sources below describe the library, but its full set of optional capabilities is broader than the configuration enabled here. Only features explicitly described as measured on this page should be treated as present in the current integration.

Three timed phases with finite requests

×

The interface allows approximately thirty seconds for a completed run, divided into separate latency, download and upload windows of about ten seconds each. A phase has a finite request configuration. The latency phase requests a set of samples, and the transfer phases use defined payload sizes and counts. The library can finish those requests before the visual phase window ends. The remaining display time is not evidence that fresh transfer traffic continues at the indicated rate throughout the whole window.

This matters especially on faster connections, where finite payloads can complete quickly. Do not estimate the data consumed by multiplying the final Mbps reading by the full visual test duration. The result is a summary of actual measured requests, while the timer manages the user interface and phase deadline. If a required phase does not return a finite positive measurement, the run cannot be treated as a complete verified result. Cancellation stops the active measurement workflow and detaches its updates, so a dismissed or interrupted run should not later appear as a newly completed test. Keeping the browser visible and the device awake reduces the chance that unrelated scheduling or suspension interferes with the observation.

Download measurement and displayed units

×

Download measures data travelling from the measurement endpoint toward your browser. The interface uses the library's throughput summary and converts bits per second into megabits per second, displayed as Mbps. A bit and a byte are different units: eight bits make a byte. A file download manager may display megabytes per second, so its number cannot be compared directly with the speed card without accounting for the unit. File size conventions and transfer overhead add further differences.

For a hypothetical illustration, a sustained transfer of eighty megabits per second corresponds to an ideal ten megabytes per second before overhead and other limits. That arithmetic does not predict the exact duration of a real download. A file server can impose a limit, a transfer can pause, a browser can process data, and your local storage may become relevant. The displayed throughput also summarizes particular requests rather than promising that every future transfer will maintain that rate. When comparing results, use the same unit, preserve the measurement method and record the context. A large number is evidence of capacity observed on that path, not proof that every small webpage operation will respond quickly.

Upload measurement and direction of a bottleneck

×

Upload measures data travelling from the browser toward the endpoint. It deserves separate attention because connections can have different capacity in each direction. Sending a large attachment, synchronizing a photo library, providing video to a call and backing up a device all depend on outbound traffic. A high download result does not establish that those tasks have enough upload capacity, and a slow upload does not necessarily imply an equally slow download.

The upload phase also uses real payloads and the library's summary. Consider other traffic when interpreting it: a background backup may share the same outbound link, while an application can wait for processing after its network transfer has finished. A useful troubleshooting observation separates the time spent sending data from the time spent waiting for an application to finish its own work. If outbound performance is the concern, record an upload measurement and the relevant application's behavior under comparable conditions. In a hypothetical comparison, a connection that delivers a much larger download figure than upload figure might handle streamed viewing comfortably while struggling with simultaneous outbound tasks. That example illustrates a direction mismatch, not a rating for a particular plan or household.

Latency, jitter and missing metrics

×

Ping and jitter in the full speed result come from the library's latency summaries and are displayed in milliseconds. Latency describes delay for the measurement exchange, while jitter describes variation among the relevant delay observations. Lower delay often helps an interactive task respond, and less variation can make its timing more predictable. Neither number alone measures how a particular application conceals delays, buffers media or recovers from disruption.

A missing jitter measurement stays missing. It must not be interpreted as zero variation, and it should not be replaced with a value inferred from download speed. The current full test configuration does not measure packet loss or latency under download or upload load. A quiet connection with a low displayed ping can still perform poorly when another task fills a queue, but this result does not quantify that loaded condition. Likewise, a pause in a call cannot be assigned a packet loss percentage from these cards. Treat each metric according to what was actually measured. If a field is absent, state that limitation explicitly in a report instead of converting the absence into a favorable score or hiding it inside an overall average.

The separate ping tool uses another method

×

The dedicated Ping Test page times twelve HTTP requests to this website. It measures the interval around the request and response handling in the browser. The destination and method differ from the full Cloudflare speed test, and the result is not ICMP ping. Browser timing can reflect connection setup or reuse, request processing and local scheduling alongside the network path. These characteristics are part of the observation and should be named when comparing it with a command line tool or a game display.

The page reports an arithmetic average, a median and average absolute change between successive samples as jitter. To calculate the median, values are sorted; an odd sample count uses the middle value, while an even count uses the mean of the middle pair. The jitter calculation preserves sample order because consecutive changes are what it summarizes. A hypothetical series that alternates between short and long responses can therefore have substantial jitter even if its average resembles a more consistent series. Twelve observations are a brief sample, not continuous monitoring. A failed request or incomplete sequence should be described as such rather than silently treated as a valid packet loss measurement.

What a completed result means in local history

×

History retains completed results that satisfy the site's stored measurement checks. Those checks verify the shape and required numeric properties of the record; they do not independently certify the network provider or prove that a measurement took place under laboratory conditions. The interface may use the word verified to distinguish real completed records from fabricated display examples. Read that label within this limited meaning, and retain the method and context when exporting a result.

The local history can contain records from the current Cloudflare method and eligible older M-Lab measurements. Their methods and destinations should not be assumed identical merely because they appear in the same table. Up to fifty completed results are retained in the browser. The history belongs to that browser profile, so another device or browser may have a different series. If storage is blocked, a measured result can remain visible without being durably saved. If the history is cleared, the site does not promise to recover it. These storage facts affect what a trend represents: it is the set of available observations in that context, rather than a complete archive of every test ever performed by the household.

Statistics and context matching

×

The Statistics dashboard calculates average download, average upload, best download and median ping from the valid local records available to it. Its download chart shows a recent subset of history. These summaries are descriptive arithmetic, not a model that removes differences in device, provider or network conditions. If you deliberately test several networks in the same browser, the general dashboard can contain that mixture. Read the individual rows before interpreting a single summary as the condition of one connection.

The latest result insights use a narrower comparison when enough metadata is available. Comparable records must match provider, country, selected network mode, local connection label, server and method, with the area label matching as well. The comparison considers recent records within a day, avoids duplicate identifiers and uses a limited recent set. Missing metadata prevents a confident match. Even this matching cannot know every environmental detail, such as whether a microwave was active or a household member started a backup. It makes the comparison more controlled but does not establish causation. A change in the median of comparable observations is a reason to investigate the circumstances, not a diagnosis of a particular cable, router or provider.

Provider contributions and acceptance checks

×

After a successful home page test, the browser automatically attempts to submit a result for provider comparisons. The server checks the country and selected mode, recognized provider identity, measurement method, result identifier and accepted metric ranges. It rejects duplicate identifiers and limits rapid repeat submissions using a fingerprint derived from request information. A locally completed test can therefore remain in Reports while its provider contribution is not accepted.

These checks reduce obvious errors and some repeated submissions, but they do not make a voluntary browser sample statistically representative. They do not independently prove the subscriber's plan, the physical connection technology or the conditions under which the device was tested. Provider names are matched to a directory, which can require normalization of network organization names and brand names. Networks absent from the directory may not produce a comparison row even when a local speed measurement succeeds. An acceptance failure should be understood in that context. Repeatedly running tests to force a public average to change is not a sound sampling method and can be limited by the service. Record a useful local observation and investigate an unexpected provider label separately.

Country averages, sample counts and time windows

×

Public provider rows use arithmetic averages of accepted submissions from the last ninety days, grouped by country, provider and the broadband or mobile mode chosen by visitors. Each accepted measurement is an observation; the count does not identify a count of independently selected subscribers, addresses or households. Several accepted observations over time can come from the same connection. A larger sample is informative about more submitted measurements, but its size alone does not cure selection bias.

Moving the map, selecting a point or searching for a city does not generate city level speed measurements. The provider comparison remains a country and selected mode view. A directory entry without samples has no measured average, and an absence of measurements is not a zero speed rating. The latest update time helps describe freshness, while the rolling window limits which records affect the calculation. Older stored submissions may remain outside that window; the window is not a retention promise. Historical aggregates can mix eligible M-Lab and current Cloudflare methods. A careful comparison therefore reads the source, time window, sample count and grouping together, and does not present the table as a census of all customers or a forecast for a specific address.

Provider directories and population information

×

Provider directory entries supply names and network metadata for the countries represented. APNIC population estimates are used where available, with other verified network metadata supporting directory coverage. Such information helps identify major networks; it is not a collection of InternetSpeedGeo speed samples. A provider can be present in the directory while its measured speed columns are unavailable. Popularity, estimated user share and measured throughput answer different questions.

Do not treat row order, a recognizable logo or an estimated user population as a recommendation about the best plan for your home. Local availability, service terms, access technology and your own requirements remain relevant, and they are not established by directory membership. A provider's public brand may differ from the organization associated with its address space, particularly with wholesale arrangements, mobile services or corporate connections. If the displayed identity seems wrong, report the country and label with supporting official information. Avoid assuming that a similar name identifies the same legal operator everywhere. The directory is contextual material for reading a measurement, and corrections to that context should be distinguished from changes to the numerical observations themselves.

Preparing a comparison that can answer a question

×

Write down the question before changing anything. For coverage, compare the same device near the access point and in the problem room. For timing, compare similar sessions at different parts of the day. For a VPN, compare authorized configurations with and without it while recognizing that the destination path changes. Record which condition changed and which conditions remained comparable. A sequence in which the device, room, network and browser all change may reveal a difference without explaining its source.

Pause avoidable transfers when you want an unloaded baseline, while respecting other people using the connection. Keep the test page active, prevent the device from sleeping and allow an earlier test to finish before starting another. Record whether you used Ethernet, Wi-Fi or mobile data, and note the selected network mode separately. A phone using Wi-Fi is still using its current fixed or shared upstream connection, while a laptop tethered to a phone can be using mobile access. The physical device category does not settle that distinction. If you are testing ordinary busy household conditions instead of a quiet baseline, state that purpose and describe the traffic present. Both observations can be useful when their different questions are clear.

Sources of uncertainty and failure

×

Precision in presentation should also be distinguished from accuracy in interpretation. A decimal place gives a formatted summary more detail, but it does not establish that a tiny change is meaningful. If two hypothetical download readings differ slightly while the room, radio traffic and route may vary, declaring a winner from that difference would be premature. Look for a pattern that persists under comparable conditions and is large enough to matter to the task you are investigating. Preserve the unrounded export when you need arithmetic, but explain the result in ordinary language appropriate to the evidence. Apparent precision can otherwise distract from much larger uncontrolled differences between sessions. The same caution applies to comparing a country average with an individual result, because the populations and measurement conditions differ even when the units match.

A browser shares device resources with other applications. Heavy processor work, power management, background tabs, extensions or restricted network settings can alter the conditions in which requests are made and timed. Wireless signal and interference, shared mobile capacity, router behavior and destination routing add more possible variation. These factors do not mean every result is useless; they explain why a surprising result needs a repeatable comparison before a cause is assigned.

When a test fails, preserve the stage and error rather than interpreting a partial speed as a complete run. Check whether ordinary websites work, whether the measurement destination is blocked by a permitted network policy, and whether the device stayed connected. A failure only on one managed network may require its administrator's explanation. Do not disable required security controls merely to obtain a number. If a result changes after a browser restart, that is evidence about the changed conditions, not proof of a specific browser defect. In a report, separate what you observed from what you suspect. A concise statement that the upload phase failed twice under a named configuration is more useful than claiming the provider has a network wide upload outage.

Data usage, privacy and repeat frequency

×

The test consumes real download and upload data. Its finite request sets and timing rules mean the amount varies with execution, so a fixed per test allowance should not be inferred from the timer. On a metered plan or a connection with limited credit, check the relevant provider usage information before running a series. Space tests around a specific question rather than running continuous loops, which can compete with ordinary traffic and distort the conditions you are trying to observe.

The privacy policy explains local history, automatic contributions and external requests in more detail. Cloudflare receives measurement requests; the home and IP tools use IP related lookups; map features request map data; optional analytics and advertising have separate consent behavior. The library's optional measurement and AIM logging endpoints are disabled in this integration, but operational processing by Cloudflare still matters. Clearing local history does not delete a contribution already sent to the server. When sharing evidence, provide the smallest relevant extract and remove unrelated account or location details. The purpose of a troubleshooting record is to preserve measurement context, not to collect every identifier available on the device.

Drawing a conclusion and reporting a correction

×

A defensible conclusion names the measured condition and the limit of the evidence. For a hypothetical example, repeated wired tests might be consistently faster than tests from a distant room on the same device. That supports investigating the wireless segment; it does not identify a particular radio setting without further comparison. Another hypothetical series might show similar speed measurements while only a specific application remains slow. That shifts attention toward the application path or service without proving the local connection is perfect.

When reporting a suspected methodology error, include the page URL, language, disputed statement and the implementation or primary documentation that supports the correction. General support can be reached at hello@internetspeedgeo.com. A helpful message distinguishes a wrong explanation from an unexpected individual measurement, because those require different checks. The review date records when this explanation was checked against the described implementation and sources. It is not an external audit date, a laboratory accreditation or a certification of every submitted result. Continue to preserve uncertainty when quoting the site: a dated browser observation with a known destination and method is useful evidence precisely because its scope can be stated clearly.

References and further reading

×

Overview

What does your speed test actually measure?

  1. Your device
  2. Wi-Fi / Ethernet
  3. Your provider
  4. Test server
One test measures this complete route at one moment. A slow result alone cannot identify which part caused it.

Understand the current test configuration, the separate HTTP ping tool, stored history and country provider averages before drawing conclusions from a result.

ContactMethodologyWiki

Useful, at a glance

Relevant next steps to help you put this page into practice.