Reliability
Share of requests that passed, each time slot weighted equally, per target
100 × yours ÷ leader More unique IPs than any network on the planet, in the world's largest residential pool. That's Shifter.More unique IPs than any network, in the world's largest pool. That's Shifter.
See the numbersAtlas methodology
Every figure in Atlas comes from one collector, one written target suite and one published scoring rule. Nothing is hand-tuned, and Shifter is measured the same way as everyone else.
Ranked by Atlas Score across 18 networks
Shifter94.1#1
Oxylabs83.4#2
Decodo81.5#3
FleetProxy80.6#4
Geonode80.3#5
DataImpulse80.1#6
ProxyEmpire79.9#7
Webshare79.3#8
NodeMaven77.7#9
SOAX77.5#10
Proxy-Seller77.2#11
Byteful75.6#12
Infatica74.6#13
Massive74.2#14
Evomi73.1#15
Rayobyte71.5#16
Proxy-Cheap70.5#17
IPRoyal66.4#18
Shifter94.1#1
Oxylabs83.4#2
Decodo81.5#3
FleetProxy80.6#4
Geonode80.3#5
DataImpulse80.1#6
ProxyEmpire79.9#7
Webshare79.3#8
NodeMaven77.7#9
SOAX77.5#10
Proxy-Seller77.2#11
Byteful75.6#12
Infatica74.6#13
Massive74.2#14
Evomi73.1#15
Rayobyte71.5#16
Proxy-Cheap70.5#17
IPRoyal66.4#18Every number in Atlas comes from the same path. These are its six stops.
88131d128c2b Every request counts for every metric, and every time slot weighs the same. Every sub-score is a ratio to the best network in Atlas (100 = the leader). Target scoring: 80% success + 20% speed under load, each against the leader. A network that is down is data, not a gap. Overall scores require complete, fresh inputs; missing values remain unavailable.
Verify it yourself: the command prints the version.
curl -s https://atlas.shifter.io/v1/methodology/88131d128c2bd29076cfbe3a83921cf29d5a95a03b34c3b226716e698eecccb1.json | shasum -a 256 88131d128c2bd29076cfbe3a83921cf29d5a95a03b34c3b226716e698eecccb1
Methodology version: 88131d128c2b
Atlas measures residential proxies: rotating and sticky, each tested as customers buy it.
A new exit IP per request or per short session, drawn from a shared pool.
A residential exit held for an advertised TTL, for multi-step sessions.
All countries with usable coverage · 237 countries tested
This is the product most buyers are actually choosing between, and the one where claims are least checkable. It gets the full suite: every target, continuous exit-IP sampling, churn tracking and the cross-network overlap audit.
All countries with usable coverage
Same suite as rotating, plus TTL verification: we hold a session for its full advertised duration and record whether the exit IP survived. Advertised TTLs are ceilings, and the gap is never disclosed.
A 200 is not a success. Each response is checked against a written rule for its target.
One live request per target through the proxy gateway.
Exit IP mapped to country, city, ASN and IP type.
Must pass a written check. HTTP 200 is not success.
The response met the written condition for that target. · success
A CAPTCHA came back as HTTP 200 instead of the page. · failure
The target refused: 403, 429 or a reset at the edge. · failure
No usable response within the target's time budget. · failure
Success, latency and IP pool, scored against the leader.
Changes logged, full dataset published as open JSON.
A collector node opens a session against the provider gateway with our own credentials, requests one target, and records the full response: status, headers, timing, body assertions and the exit IP the target saw.
The exit IP is looked up against public registry data for country, state, city and ASN, then classified as eyeball, hosting, mobile or business. Nothing is taken from the provider dashboard.
A 200 response is not a success. Each target has a written pass condition, such as a parsed element, a JSON shape or a token, and a response that fails it is recorded as a challenge, a hard block or a timeout.
The run is diffed against the previous one, changes above the reporting threshold are recorded, and the whole dataset is published as a single JSON file alongside Atlas.
Only a pass is a success; the other three are the failures that vendor dashboards report as traffic served.
The response satisfied the written condition for that target: the parsed element, the JSON shape or the token is present in the body we got back.
A CAPTCHA or interstitial was served instead of the page. It arrives as HTTP 200, which is how success figures based on status codes end up looking almost perfect.
The target refused outright: 403, 429, a block page, or a connection reset at the edge before any content was served. Still counted as traffic.
No usable response inside the budget for that target, including errors raised by the proxy gateway itself rather than by the site.
exitIp country state city asn asnName kind status ttfb / total bytes sessionId target vantage Each request adds to every figure below, grouped by the question a buyer asks.
Every figure can be read over three time ranges
The freshest view: today's defences on every target.
Adds repeat IPs, which take more than one day to see.
The range the Atlas Score and ranking use.
New data every hour · 46,293,653 requests in the last 24 hours · 146,064,199 requests in the last 4 days · 146,064,199 requests in the last 4 days
Distinct exit addresses actually observed answering our requests.
Advertised pool size describes inventory. It is unverifiable by construction.
Share of exits in the window never seen in any previous window.
Churn determines whether your session survives, and nobody publishes it.
Share of distinct exits in the 7-day or 30-day window that appeared on two or more days.
A rotating residential pool should keep this low. A high figure is inventory that behaves like static IPs.
Mean consecutive requests one address served before the gateway rotated us away.
It is the difference between a sticky product and a rotating one, regardless of marketing.
Share of sticky sessions that survived their advertised TTL.
Advertised TTLs are ceilings, not guarantees, and the gap is never disclosed.
Share of the network's unique IPs also seen in another network's pool; the rest is its own (exclusiveShare). The network holding most of it is named as the likely upstream (likelyUpstream).
Reselling is normal. Paying a markup for the same addresses without knowing it is not.
Unique IPs seen in the window as a share of the pool size the network advertises.
An advertised pool is a claim. This is how much of it actually shows up.
A 1 to 10 score per target, from the share of requests that met the written pass condition. The best network on each target reads 10; the others keep their measured distance.
Every request counts (census, pulse and load); each time slot weighs the same, so a busy census hour cannot drown out a quiet night. A network's overall score is the mean of its target scores, each weighted by the target's difficulty, 1 to 5.
A single success figure is usually quoted against an undefended endpoint, where nearly every request succeeds on every network.
Time from request start to last byte on normal requests (the hourly pulse), through the proxy, as p50, p90 and p99, on the plain-HTTP target only. The latency under heavy load is shown beside it.
Vendors quote connect time on an idle connection, which excludes both the part you wait for and the slow-down you hit at volume.
A 1 to 10 score per target, ranked on the share of requests served a CAPTCHA or interstitial: fewer challenges score higher. A challenge counts as a failure.
A challenge returns HTTP 200, so any success figure based on status codes silently counts it as a win.
Outage time per country tested, so an outage in one of two countries costs half.
1 − downtime ÷ (window × countries). An outage opens when a pulse finds the network down and closes with the next healthy pulse; a census the network spoiled is rerun once it is back.
Proxy outages are almost always partial, so binary uptime always reads 100%.
Minutes of measured outage in the window, and the outages opened in it (incidentCount), each with its severity, reach and length.
Status pages report what the provider admits. These are the outages the collector hit.
Share of unique exit IPs whose IP database usage type is hosting or datacenter, among IPs with a usage type, on networks sold as residential.
This is the single most misrepresented property in the market.
Unique exit IPs on hosting or datacenter networks: the count behind the DC ASN share.
A share can hide a large number on a large pool.
Share of unique exit IPs that the IP database places in the country the requests asked for, among IPs with a known country.
A pool sold by country is only worth what it delivers in that country.
Share of sampled exits appearing on public reputation lists. Shown where measured; not part of the score.
It predicts failures you will blame on your own code.
DNS, WebRTC, proxy headers, TLS fingerprint and IPv6 fallback, asserted from the target side.
These are sold as features and almost never verified.
Distinct countries, first-level subdivisions, cities and autonomous systems among the observed exits. Higher is better on every count.
Advertised "195 countries" is a catalogue claim. These totals are what the collector actually saw.
$ / GB is the price of the first plan of 200 GB or more (the largest plan when none is that big), with the range up to the entry plan.
Value is the network's place on cost per 1,000 successes among the priced networks: 100 for the cheapest, falling evenly to the last.
The entry rate is what a trial costs, not what a buyer at volume pays.
Price normalised by the share of requests that passed and by payload, across all eight targets.
On each target: (1,000 × measured data per request) × lowest $ / GB ÷ the share of requests that passed, so failed requests are paid for. The eight targets are averaged with equal weight.
Rate cards price bandwidth. Nobody buys bandwidth.
Requests one success takes over the window: requests ÷ successes.
Every failed attempt is billed bandwidth. This is the retry tax a rate card leaves out.
Every figure can be read over the last 24 hours, the last 7 days, or the last 30 days. The collector writes a bucket every 5 minutes, close to real time on the pools. The longer windows are that same stream looking further back.
Requests are spread evenly across the window and across targets, so a network cannot look better by being fast at 04:00. Session and rotation settings are left at each provider's defaults; we measure the product as sold, not a tuned configuration.
Counted across the selected window.
Distinct exits seen in that window: the last 24 hours, the last 7 days, or the last 30 days.
Share of unique exits seen on two or more days, over the last 7 or 30 days. Not shown on 24 hours, which cannot show a multi-day repeat.
List price stays the same; success changes with the window, so cost per delivered result is restated for each one.
Outage time over the selected window: 1 − downtime ÷ (window × countries tested). A network that is down is data, not a gap: its slot counts as filled and down, lowers uptime and opens an incident. Last 24 hours, last 7 days, or last 30 days.
Each target's scores follow the selected window, as every other figure does.
The window control on the Overview switches every figure at once. Where a figure is only measured over 24 hours, it stays on 24h rather than moving.
The measurement we have not seen published anywhere else, and the one most likely to change a purchasing decision.
Because every network is tested in the same window with the same requests, we hold the actual exit addresses each one handed us. Intersecting those sets answers a question the market never addresses: how much of a pool is genuinely that provider's inventory, and how much is the same addresses you can buy elsewhere.
Shared counts are symmetric between a pair. The two percentages are not, because pool sizes differ, and the asymmetry is the finding. A small pool mostly contained inside a much larger one, where the reverse is barely true, is resold inventory.
Both networks tested concurrently. Pools rotate constantly, so sampling two providers on different days produces a low overlap that means nothing.
The exit address is taken from what the target observed, never from a provider dashboard.
Only address-level matches count. Two networks sourcing from the same ISP will share an ASN while sharing no addresses, and conflating the two would be a serious error.
Overlap is reported as a floor. We intersect what we saw; two pools may share more than that, never less.
Each part is compared with the best network in Atlas, weighted, and added up.
Atlas Score = Σ(sub-score × weight) ÷ Σ(weight)
Each part is yours divided by the best network of your type. The leader scores 100. 100 × yours ÷ best
0 to 100
The weighted mean of the eight parts, shown with a grade.
These are the default weights. The leaderboard lets you re-weight the mix for your own workload.
Every part is measured against the best network in Atlas. There are no fixed targets to hit.
Every network in Atlas is a residential proxy network, tested on the same targets with the same requests, so every score compares like with like.
On every part, the best network in Atlas scores 100 and everyone else scores their share of it. Half the leader's pool is 50; twice the leader's latency is 50.
Counts (IPs, countries, states, cities, ASNs) are daily counts averaged over the days a network has data, so one day of testing compares fairly with thirty. Rates such as success, latency, availability and cost use the whole window.
The headline score covers the last 30 days, or fewer for a network that is newer. Leaders are found again every time scores are calculated, a few minutes after each test batch, so the 100 moves when a network gets better.
Each one reads a single measurement and sets it against the best network in Atlas, so 100 always means the leader. The tag says which direction is better.
Share of requests that passed, each time slot weighted equally, per target
100 × yours ÷ leader Median latency of normal requests (the hourly pulse), p50, on the HTTP target
100 × leader ÷ yours Each light-load (100-request) latency × the slow-down its runner region measured at census concurrency in its fastest country. HTTPS and other targets carry no latency.
Capacity-weighted availability over the window
100 × yours ÷ leader Unique exit IPs answering requests, averaged per day
100 × yours ÷ leader Live unique exits, not the advertised pool size.
Countries, states, cities and ASNs seen, averaged per day
0.30·r(countries) + 0.20·r(states) + 0.25·r(cities) + 0.25·r(ASNs) r(·) = 100 × yours ÷ leader on that count.
Clean share: 1 − share of exit IPs whose IP database usage type is hosting or datacenter
100 × yours ÷ leader Hosting IPs in a residential pool lower this part.
100 × share of exit IPs located (IP database) in the country the requests asked for
100 × yours ÷ leader Per unique exit IP; IPs without a known country are left out.
Lowest list price per GB of the pool tested
100 × leader ÷ yours Each figure over the last 30 days, with the sub-score it earns below, coloured by its place
| Network | Success weight 22 | Latency p50 weight 14 | Availability weight 10 | Unique IPs weight 12 | Cities · ASNs weight 8 | DC ASNs weight 10 | Authenticity weight 10 | Cost / 1k weight 14 | Atlas Score |
|---|---|---|---|---|---|---|---|---|---|
1 Shifter | 9.4/10 99.3 | 486 ms 71.2 | 100.00% 100.0 | 8,948,662 100.0 | 103,632 · 24,797 98.7 | 0.6% 99.5 | 100.0% 100.0 | $4.43 88.9 | 94.1 A+ |
2 Oxylabs | 9.9/10 99.9 | 480 ms 72.1 | 99.95% 100.0 | 7,167,058 62.6 | 101,067 · 22,478 96.4 | 1.3% 98.7 | 100.0% 100.0 | $10.03 44.4 | 83.4 A |
3 Decodo | 10.0/10 100.0 | 504 ms 68.7 | 99.97% 100.0 | 7,032,104 63.6 | 100,995 · 22,414 96.7 | 1.3% 98.7 | 100.0% 100.0 | $11.38 33.3 | 81.5 A |
4 FleetProxy | 6.8/10 97.2 | 919 ms 37.6 | 99.82% 99.8 | 7,150,983 59.2 | 81,755 · 18,319 84.8 | 0.7% 99.3 | 100.0% 100.0 | $9.07 72.2 | 80.6 A |
5 Geonode | 9.5/10 99.1 | 810 ms 42.7 | 99.95% 100.0 | 3,286,685 25.3 | 55,566 · 17,508 78.3 | 0.0% 100.0 | 100.0% 100.0 | $3.30 94.4 | 80.3 A |
6 DataImpulse | 6.3/10 98.9 | 471 ms 73.5 | 99.73% 99.7 | 880,984 10.7 | 43,269 · 9,932 63.9 | 0.0% 100.0 | 100.0% 100.0 | $4.45 83.3 | 80.1 A |
7 ProxyEmpire | 7.6/10 97.5 | 799 ms 43.3 | 99.78% 99.8 | 6,036,035 54.9 | 92,380 · 20,256 91.2 | 0.5% 99.5 | 100.0% 100.0 | $9.54 61.1 | 79.9 B+ |
8 Webshare | 8.8/10 99.6 | 770 ms 44.9 | 100.00% 100.0 | 3,759,734 40.1 | 84,722 · 18,356 87.6 | 1.0% 99.0 | 100.0% 100.0 | $9.44 66.7 | 79.3 B+ |
9 NodeMaven | 7.3/10 97.9 | 512 ms 67.6 | 99.32% 99.3 | 5,356,729 51.9 | 97,252 · 21,120 94.2 | 0.6% 99.4 | 100.0% 100.0 | $16.01 22.2 | 77.7 B+ |
10 SOAX | 3.1/10 99.3 | 418 ms 82.8 | 100.00% 100.0 | 1,244,359 12.4 | 47,634 · 13,493 71.0 | 1.3% 98.7 | 100.0% 100.0 | $9.75 50.0 | 77.5 B+ |
11 Proxy-Seller | 6.4/10 97.1 | 753 ms 46.0 | 100.00% 100.0 | 3,778,210 41.1 | 79,584 · 17,490 84.9 | 0.7% 99.3 | 100.0% 100.0 | $9.69 55.6 | 77.2 B+ |
12 Byteful | 9.7/10 99.8 | 517 ms 66.9 | 100.00% 100.0 | 3,884,045 36.3 | 65,731 · 12,042 57.9 | 1.3% 98.7 | 100.0% 100.0 | $10.84 38.9 | 75.6 B+ |
13 Infatica | 6.5/10 97.6 | 709 ms 48.8 | 99.91% 99.9 | 4,818,619 47.1 | 80,869 · 18,151 86.0 | 0.8% 99.3 | 100.0% 100.0 | $13.91 27.8 | 74.6 B+ |
14 Massive | 7.0/10 99.1 | 346 ms 100.0 | 99.95% 100.0 | 1,984,735 17.9 | 67,363 · 12,363 59.2 | 0.1% 100.0 | 100.0% 100.0 | $23.52 11.1 | 74.2 B+ |
15 Evomi | 8.6/10 99.0 | 384 ms 90.1 | 99.80% 99.8 | 867,174 10.7 | 40,903 · 10,855 63.4 | 0.0% 100.0 | 100.0% 100.0 | $20.07 16.7 | 73.1 B+ |
16 Rayobyte | 1.0/10 77.5 | 888 ms 39.0 | 99.94% 99.9 | 2,101,773 16.5 | 60,854 · 17,869 79.5 | 2.4% 97.6 | 100.0% 100.0 | $7.49 77.8 | 71.5 B |
17 Proxy-Cheap | 1.0/10 75.6 | 1.02 s 33.9 | 92.34% 92.3 | 562,185 9.2 | 35,760 · 7,408 60.6 | 0.6% 99.4 | 100.0% 100.0 | $2.73 100.0 | 70.5 B |
18 IPRoyal | 7.2/10 98.5 | 1.08 s 32.1 | 99.92% 99.9 | 2,892,443 26.0 | 66,362 · 17,149 81.7 | 1.6% 98.4 | 100.0% 100.0 | $25.40 5.6 | 66.4 B |
Six honest gaps, so you know when to test for yourself.
Latency is measured from our nodes, so yours will differ.
We count the IPs that answered, not the inventory a provider claims.
We buy a mid-tier plan. A different contract can perform differently.
A success score describes one week, not the network forever.
Low, rate-limited traffic says little about very large scraping loads.
Pool overlap is published in full but never folded into the score.
Stating limitations is not a disclaimer, it is part of the method. Every one of these is a reason your numbers may legitimately differ from ours.
Latency is measured from our nodes, not yours. A network that looks slow from Frankfurt may be fast from São Paulo. Region is recorded per request and the collector runs from eight regions, but your own numbers will differ and that is expected.
We report distinct exit IPs we actually saw in a window. A provider claiming 90 million addresses is not necessarily lying; they are describing inventory, while we describe what answered. Both numbers are useful and only one can be checked.
Some networks route paying tiers through better inventory. We buy a mid-tier plan and say so. If your contract differs, test it yourself and compare.
Anti-bot vendors ship changes weekly. A success score is a measurement of one week, not a property of a network. Trends matter more than single numbers.
Request rates are deliberately low and rate-limited per target. This keeps the test ethical, and it also means we do not measure how a network behaves under the load of a large scraping operation.
Pool overlap changes decisions about redundancy and effective pool size rather than describing quality, so it is published in full and folded into nothing. An Atlas Score that absorbed it would obscure both.
Left out on purpose
Customer reviews are published beside the data and reach the Atlas Score nowhere.
A written review captures things no measurement can reach: whether support answers, whether billing is honest, whether the documentation is usable. That is genuinely valuable, and it is why reviews are published next to the measurements.
It is also unreproducible, which is the one property the Atlas Score has to have. Averaging an opinion into a measurement produces a number that is neither, so the two sit side by side and are never mixed. A five-star network with a bad score is telling you something specific, and blending them would erase it.
A verified review requires proof of purchase: a billing document showing the provider name and a month. We check those two facts, publish the month only, and delete the file.
Unverified reviews are published and labelled, and excluded from the verified average.
The data and the formula are public. Three ways to check.
Download the exact dataset Atlas was built from and recompute any number on it.
Apply the published formula to the dataset and you get the same ranking.
Run our open-source tests on your own account with the AI coding agent you already use, then compare.
The strongest is the third: test a network on your own account and compare your figures with ours. If the shapes match and only the absolute numbers differ, the method is sound and the difference is your vantage point.
Every rule on this page is public. Order with a 3-day money-back guarantee and test Shifter with the same targets, checks and scoring.