A headline success rate rarely names the site it was measured on. When it does, the site is usually an undefended test page, where almost any proxy network looks excellent. That number says little about how the same network does on a search engine, a marketplace or an AI assistant.
If you are choosing residential proxies for real work, the useful question is not “what is this network’s success rate” but “what is its success rate on sites like mine”. This guide explains why the answer changes so much from site to site, and how the Atlas Targets page is built to answer it.
One number hides the spread
Picture two networks. On a plain page with no defences, both pass almost every request. On a search engine, one keeps passing most requests while the other is challenged on a large share. Average the two targets together and both look respectable, and you have lost the only result that would have told them apart.
This is the normal pattern, not an edge case. Networks tend to look alike on easy pages and spread apart as the defences get harder. A single blended rate is mostly a measure of how easy the test targets were.
What different sites actually judge
Websites do not all run the same check. Each kind of defence looks at a different property of the request, so each one exposes a different weakness:
- Undefended pages judge nothing. A failure there is the proxy network itself: a dropped connection, a slow gateway, a timeout.
- Marketplaces apply moderate bot defences to price and stock pages, the most common scraping job.
- Search engines challenge addresses that many customers share, so they expose pools that are used too hard.
- Public profile pages refuse known proxy ranges and hosting networks, which tests how residential a pool really is.
- App endpoints check the connection as well as the address, catching exits that only look residential.
- Managed challenges score the exit address’s reputation before serving anything.
- High-contention queues defend hardest of all and admit only clean, rarely used exits.
Our anti-bot stack lookup shows how to find out which of these protects a site before you build anything against it.
Easy pages test the network, hard pages test the pool
A useful way to read per-target results is to split them in two.
The baseline targets tell you about the network: its gateway, its routing and its uptime. If a network fails an undefended page, nothing about the target explains it.
The defended targets tell you about the pool: how clean the addresses are, how many customers share them, and whether they are really home connections. This is where most buyers spend their money, and where networks differ most.
The gap between the two is the most informative figure on the page. A network that is near the top on the baseline and falls hard on defended targets has a reliable gateway and a tired pool.
The reason defended sites separate networks so sharply is that they judge the address, not just the request. A genuine residential proxy exit is a home connection with a consumer ISP behind it, and most defences give that kind of address the benefit of the doubt. An address that is shared by many customers, recycled quickly or registered to a hosting company loses that benefit. On an undefended page none of this shows. On a search engine or a managed challenge it decides most of the result.
Rankings change by target
Because each target judges something different, the order of networks is not fixed. A network can lead on marketplace pages and trail on search, or the reverse. If your workload is mostly one kind of site, the overall ranking matters less than the column for that kind of site.
The comparison by target puts two networks side by side on every target, which makes these reversals easy to spot.
Why per-target results use a short window
Sites change their defences often, sometimes several times a week. A month-long average on one target would mix several different defence states into one number that describes none of them.
That is why per-target success on Atlas is reported over the last 24 hours, while the overall ranking uses a longer window. The methodology on time windows explains which figures use which range and why.
How to read the Targets page
The Targets page lists every target from easiest to hardest, with the written pass rule for each, and then shows every network’s success on every target in one table. To use it:
- Find the target closest to your own. Match the kind of defence, not the brand of the site.
- Read that column first. Treat the overall average as context, not as the answer.
- Compare against the baseline. A large drop from the plain page to your target points at the pool.
- Check the pass rule. Make sure it tests for the same kind of evidence you need.
- Treat N/A as untested. A network that has not been sent to a target yet is not scored as zero, and should not be read as a failure or a pass.
Testing on your own site
No measured target is your target. Use Atlas to build a shortlist, then confirm on your own work:
- Send each candidate the same requests, from the same locations, at the same time.
- Judge every response with a written pass rule, not a status code.
- Run long enough to catch more than one defence state, at least a few days.
- Report your own target separately from any easy control page.
If the kind of site you depend on is missing from the suite, suggest it from the list of upcoming targets at the end of the Targets page. A target earns a place when its pass rule can be checked by code and its result is not already predicted by a target in the suite. If you have already measured it yourself, you can share your own test results with Atlas.