Independent registry

Methodology

Last updated

How ProxyBench tests proxies

Methodology version 0.6.2 · updated 2026-11-21

ProxyBench buys proxy access, sends requests through it from its own measurement servers to its own control server, and publishes the results with the sample size next to every figure. Providers cannot edit the numbers. Anything a provider says about itself (price, countries, pool size, trial terms) is labelled as claimed.

What we measure in one sentence

A test run is 100 to 200 requests sent through one proxy product from one measurement server to our control server, which reports the address each request arrived from. Published figures are calculated over a window of the latest runs (see “Window of runs”).

How a test works

  1. A measurement server (a “node”) in a known city sends requests through the provider’s proxy.
  2. Every request goes to our control server, which answers with the exit IP address it saw.
  3. Each exit address is looked up in two independent geolocation databases and one network (ASN) database.
  4. The requests of one run are reduced to a few figures: success rate, response time, and the make-up of the address pool.
  5. Each run also includes 20 header probes over plain HTTP, which check which request headers reach our control server.

All nodes run the same software version, the same settings and the same schedule. The only difference between nodes is their location.

A new address for every request

Only runs in which every request is sent with a new exit address are used in the tables. Providers switch addresses in different ways (a gateway, a range of ports, a mode set in the login), so the setup differs by provider, but the aim is the same. Sticky sessions, where one address is held for a period, are not compared.

When a request counts as successful

A request is successful only if all of the following are true:

  • the proxy accepted the connection and answered;
  • the control server returned HTTP 200 with a valid response;
  • the exit address is not the address of our measurement server (the request really went through the proxy);
  • the whole request finished within 10 seconds.

Everything else is a failure and gets a reason: timeout, connect, proxy_auth, tls, bad_status, proxy_forbidden, proxy_rate_limit, proxy_gateway, proxy_reset or other. For each request we also keep the status code the proxy returned to the connection request and the error message, when there is one.

Failed runs are never repeated to get a better number. The 10-second limit is fixed; if it ever changes, a new measurement line starts.

Expired access is not a failure. When our access to a product ends (the term runs out, or the traffic included in it is used up), the product is skipped and its last test date stays visible. When a plan has a small amount of traffic, we may stop testing it shortly before it runs out, so that an empty package is never counted as a failure of the provider.

Run size and schedule

Proxy type Requests per run Concurrent
Residential 200 5
Mobile 200 5
ISP 100 5
Datacenter 100 5

Each request uses a new address (see above). For products with a port list, each request uses a different port in random order.

Datacenter products are shown in two tables because they are sold and measured differently: static products (a list of addresses, priced per IP) and rotating pools (one gateway that changes the address, priced per GB). Their figures are never mixed in one range.

Runs start in fixed time slots. Products in the main proxy countries use 04:00, 10:00, 16:00 and 22:00 UTC. Additional proxy countries and products with short access also use 01:00, 07:00, 13:00 and 19:00 UTC, so that eight slots a day share the load of the measurement servers. Every product is first tested every day for 7 days, in different time slots. After that the schedule depends on the proxy country and on how our access is paid:

  • Germany, the United States and Japan (the main proxy countries): once a week, for as long as our access lasts.
  • Other proxy countries, access paid by traffic (pay as you go): once every two weeks.
  • Other proxy countries, access for a month: once a week.
  • Free trial or access for a few days: 7 runs within the term of the access, in different time slots, several a day if needed, and no runs after that. Where the term is too short for 7 runs, the figure rests on fewer runs, and the page shows how many runs stand behind each figure.

Runs are spread over time slots so that the measurement servers are never overloaded. A product that is added while the slots are full may start a day later.

How access is obtained

We test the access a new customer gets: a paid plan bought with our own money, or a free trial offered to everyone, on the same terms as any user. A trial can differ from a paid plan (traffic limit, pool, speed), so figures collected on a free trial apply to the trial only and are marked as such on the page; unmarked figures come from a paid plan. Access we are given by a provider on terms not open to other users is not used, or is clearly marked on its page.

Window of runs

Figures in the tables and cards are calculated over the latest runs, up to 7, not over a single run. A single run of 200 requests ranks noise: its margin is about ±1.4 percentage points at 99% success.

  • The window is built separately for each combination of provider, proxy type, region of measurement and proxy country.
  • If a combination has fewer than 7 runs, all of them are used and the page shows “N runs” next to the figure.
  • The window does not drop runs because of their age. Products that are tested less often (see “Run size and schedule”) have a window that spans a longer time, which is why the date of the latest run is always shown.
  • Success rate is the sum of successful requests divided by the sum of all requests in the window. It is not an average of the percentages of single runs, so a run with many requests weighs more.
  • TTFB p50 and p95 are calculated over all successful requests in the window.
  • Unique IPs are counted inside each run and then added up. The same address seen in two different runs is counted twice.
  • A failed run stays in the window. If a connection could not be established in a run, it is not repeated and not removed. The card says “Connection failed in K of N runs” with the date of the run.
  • Last tested is the date of the latest run.
  • Connect time is taken from the latest run.
  • Country figures (Country match and how the rest splits, see “Country check”) are counted over the same window as success rate and TTFB, from the addresses that both databases place in a country. This applies to the table of proxy countries on provider cards and to country pages. The exception is the pool check at the top of a card measured from several regions: it combines the latest run of each region (see “Regions: what is shown separately and what is combined”), so its percentages can differ from the table.
  • ASN figures (ASN count, Top 3 ASN share) on a provider card are taken from the latest run; on country pages they are counted over the window.
  • On provider cards, the country status of static ISP and static datacenter addresses is taken from the latest run.

The figures

Figure How it is calculated Published from
Success rate successful requests ÷ all requests × 100, over the window 100 requests
TTFB p50 median time to first byte of successful requests, over the window 100 requests, 50 successful
TTFB p95 time that 95% of successful requests do not exceed, over the window 200 requests
Connect p50 median time to establish the connection to the proxy, latest run 100 requests
Unique IPs distinct exit addresses among successful requests, shown as “unique / successful” Residential, Mobile and rotating Datacenter, 100 successful
Repeat IP share addresses seen more than once ÷ unique addresses × 100 same
Country figures share of addresses per country status (see below) 30 unique addresses
ASN count, Top 3 ASN share distinct networks, and the share of addresses in the three largest 50 unique addresses

Percentiles use the nearest-rank method. If there is not enough data, the figure is not shown at all. It is never shown as zero.

The sample size is printed next to each figure. With 200 requests the margin of success rate is roughly ±1.4 percentage points at 99% and ±3.8 at 92%; a window of several runs narrows it.

Reading TTFB. Time to first byte is measured to our control server in the Netherlands, so it covers the whole path of a request: from our node to the provider’s gateway, through the provider’s network to the exit address, and from that address to our control server. It is therefore higher for proxy countries that are far from the Netherlands. Compare providers inside one proxy country, measured from the same node; do not compare the TTFB of one proxy country with that of another.

Country check

Each exit address is checked against MaxMind GeoLite2 and IPinfo Lite. The country the provider claims is compared with what both databases say.

Both databases are refreshed every week, and a run is checked against the database versions available when its results are processed. A geolocation database can lag behind real changes (for example when a network block is moved to another country), so a country result reflects the databases as of the check, not the address’s registration in the provider’s panel.

Status Meaning
Confirmed by both databases both databases name the country the provider claimed
Databases disagree the two databases name different countries
Both databases contradict the claimed country both databases name the same country, and it is not the one the provider claimed

Country match is the share of addresses with the status “Confirmed by both databases”, calculated over the window of runs, and n is the number of addresses behind it (addresses are counted inside each run and added up, as for unique IPs). Country match is not the share of addresses outside the claimed country: when the two databases disagree, we cannot say that either of them is right, and the address is counted as not confirmed, not as wrong. The card shows how the rest splits between “Databases disagree” and “Both databases contradict the claimed country”.

We publish a disagreement as it is: it is what a customer of that provider gets, and it is not corrected or hidden. An address for which at least one database has no country is left out of the shares. ISP and static datacenter products are a few fixed addresses, so their country result is not a share: with a single address it is one status, and the page shows that the figure rests on a single address; shares are published from 30 unique addresses. For rotating datacenter pools, the card shows shares as for residential products, and the table shows one status for the pool, taken from the addresses whose country is known in both databases: “Confirmed by both databases” when at least half of those addresses are confirmed; otherwise “Databases disagree” if any address splits the two databases; otherwise “Both databases contradict the claimed country”.

Default order of the tables

By default tables are ranked, not sorted by the raw success rate. A provider measured on a few requests can show a high rate by chance, so the ranking uses a cautious estimate of the success rate: the lower end of a 95% confidence interval (Wilson score), calculated from the successful requests and all requests in the window. The more requests stand behind a figure, the closer the estimate is to the shown rate; a figure from one run counts for less than the same figure from seven runs.

Providers whose estimates are within 0.5 percentage points of the best provider of a group (1 percentage point for Mobile) are treated as equal in reliability. Inside such a group the order is by TTFB p50 (faster first), then by the date of the last run (newer first). The next group starts with the first provider outside the group. The ranking is calculated for the selected region.

The success rate shown in the table is not changed by the ranking: it is always the measured figure. You can re-sort any table by any column. Providers that are not measured yet are listed separately.

Regions: what is shown separately and what is combined

Success rate and response time depend on where the test is run from, so they are shown per region and are never averaged into one number. Where a provider has been measured from two or more regions, a “Measured from” switch selects the region.

Figure Shown as
Success rate, TTFB, Connect separately for each region
Unique IPs in the table for the selected region
Repeats, country figures, ASN figures in the card one figure across the addresses seen from all regions

Pool figures in the card are combined because they describe the provider’s address pool, not the route to it. An address seen from two regions is counted once. The label states how many regions and addresses stand behind the figure, for example “3 regions combined · 560 unique IPs”. A provider measured from three regions will show more unique addresses than one measured from a single region; compare the percentage shares, not the raw counts.

Two summary figures are shown only when every region has a window of runs. Both are calculated from the same window of runs as the regional figures (see “Window of runs”):

  • Combined success = successful requests from all regions ÷ all requests from all regions × 100, over the windows.
  • Spread between regions = highest regional success rate − lowest, in percentage points, over the windows.

Proxy countries and nodes

Two different places appear in our figures, and they are never mixed:

  • A node is the location of our measurement server (Europe, Asia or the USA). It is where the requests start from.
  • A proxy country is the country of the proxy address we ask the provider for, and the country we then check the exit addresses against.

Germany is the base proxy country: it is tested from every node, and regional and combined figures refer to it. Other proxy countries, such as the United States, Japan, the United Kingdom or Canada, are tested from the European node and, for some products, from the other nodes as well. Their figures say which node they come from.

On a provider card, the block “Proxy countries tested” lists residential, mobile and rotating datacenter products by country, with success rate, TTFB p50, Country match and the number of unique addresses. A row for an additional proxy country appears once the product has at least 3 runs in that country, so a country that was just added is not shown yet. Static ISP and static datacenter products are not listed in that block; their country result is shown on the country page.

Where a country has its own page, products are grouped by type there. Figures from different nodes are never joined, and a country that is measured from one node only is described as measured from that node. The block “What the data shows” on a country page is generated from the measurements, with no hand-written figures. A group of products is included only when at least 3 products in it have 100 or more requests in the window, and country figures only when at least 100 addresses stand behind them; for additional proxy countries, a product also needs at least 3 runs. Where a product is measured both in that country and in Germany, the block compares the two.

We add proxy countries over time, as access and traffic allow. A product is measured in an additional country only if the provider offers that country and the first check passes:

  • Rotating products: the exit addresses must be confirmed as that country by both databases for at least half of the addresses, and must not be contradicted by both databases for more than 30%.
  • Static ISP and datacenter addresses: at least one database must name the country, and the two databases must not both name the same other country. If the databases disagree, the address is still measured and the disagreement is published as a status.

A product that does not pass is not added, and its page does not show that country.

Replacing a measurement server

Data from servers in different locations is never joined into one line. When a node is replaced by another in the same location, both run in parallel for 7 days. The line continues only if the 7-day medians differ by no more than 2 percentage points of success rate (3 for mobile) and 15% of TTFB p50. Otherwise a new line starts and the change is marked.

Prices

Prices are not measured. They are taken from the provider’s public pages, shown as “from” with a unit (per GB, or per IP per month) and labelled with the date they were checked. Where a provider sells the same product for several terms, the price shown is for a plan of about 30 days; shorter terms are not shown, and prices are not converted between terms. Plans change often; the provider’s own page is the source to trust before buying. Tariff details are listed on the provider’s card, not in the table.

Citing our figures

Each provider card shows a ready sentence for citing its figures, with the sample size, the region and the month, and the same values as the card. Providers can use a badge “Measured by ProxyBench”. The badge states only that a product was measured: it carries no figure, date or rank, and it is not an endorsement. If a figure looks wrong, send details on the corrections page and we review it.

Limits of this methodology

  • Times and success are measured to our control server in the Netherlands. They are not the provider’s speed or success rate on any particular website, and results on protected websites can differ. Time to first byte depends on the distance between the proxy country and the control server (see “The figures”).
  • Download speed is not measured. It would be limited by the bandwidth of our measurement servers rather than by the provider, so we publish success rate and time to first byte instead.
  • A run is 100 to 200 requests. Each product is tested in fixed time slots of the day, so day and night differences are only partly visible, and a gap between two providers can partly come from the time of day of their runs.
  • Products in additional proxy countries are tested less often than in the main ones, so a window can span weeks. The date of the latest run and the number of runs are shown next to every figure.
  • Static ISP and datacenter results rest on the few addresses we buy. They describe those addresses, not a whole pool.
  • One plan per proxy type is tested. Figures apply to that plan. If it is a free trial, it may differ from paid plans.
  • Sticky sessions and changing the address on command are not tested. Port-based plans are tested the same way as plans sold by traffic.
  • IPv6 is not measured: datacenter and ISP products are tested on IPv4.
  • An ASN shows who owns a network. It does not prove that an address is a home or mobile line, and we do not claim that it does. We do not classify networks as operator or hosting.
  • A geolocation database gives a country for an address, not its physical location. Country figures say what two databases report, not where a server stands.
  • Individual exit IP addresses are never published.
  • A few nodes are a few vantage points, not the whole world.

Some links to providers are affiliate links, and ProxyBench may earn a commission when you buy through them. This is disclosed in one line above each table and on provider cards. Providers without an affiliate programme are marked “No commission”. Commission does not change how a provider is tested. By default tables are ranked as described in “Default order of the tables”, and you can re-sort them by any column.

Data sources

This product includes GeoLite2 data created by MaxMind, available from maxmind.com. IP address data powered by IPinfo. Flag icons: flag-icons (MIT License).

Version history

Any change to the method gets a new version number and date. A change that affects the figures (timeout, run size, definition of success) starts a new line in the history.

Version Date Change
0.6.2 2026-11-21 Wording only. The control server is stated to be in the Netherlands, and a paragraph on how to read time to first byte is added: it covers the whole path from our node to the control server and grows with the distance between the proxy country and the Netherlands, so providers are compared inside one proxy country, not between countries. No figure and no calculation changed
0.6.1 2026-11-20 Country match (“Confirmed by both databases”) in the table of proxy countries on provider cards is counted over the window of runs, as defined in “Country check”. Before this date that column was taken from the latest run, so some values and the number of addresses behind them changed. The section “Window of runs” now states which figures are counted over the window and which come from the latest run
0.6 2026-11-19 Schedule by proxy country and type of access: main proxy countries weekly after 7 daily runs; other countries every two weeks (pay as you go) or weekly (monthly access); free trials get 7 runs within their term. Eight time slots a day (01:00, 07:00, 13:00 and 19:00 UTC added). Window of runs does not drop runs by age. Section on proxy countries and nodes rewritten (proxy country versus node, Country match and n, first check before a country is added, rows of additional countries on cards from 3 runs, automatic “What the data shows” block on country pages). Static ISP and datacenter addresses: country result as a status. Citing our figures and the badge. Limits: geolocation databases report a country, not a location; download speed is not measured; static results describe the addresses we buy
0.5 2026-11-15 Figures over a window of up to 7 runs; failed runs stay in the window; a new address for every request; more failure reasons recorded; header probes; schedule (fixed time slots; daily for 7 days, then weekly; products with short access are tested several times a day); tested access is a paid plan or a public free trial, and free-trial figures are marked on the page; prices (about 30 days) and affiliate disclosure described; databases disagreement published as it is; default ranking of tables by a cautious estimate of success rate, then speed; rotating datacenter pools are measured and shown in their own table, with unique addresses and the country check counted as for residential products; geolocation databases refreshed weekly. The word “run” is used throughout
0.4 2026-10-03 Regional figures, combined pool figures, exit countries, node replacement rule
0.2 2026 First public description