← all guides

Subnet diversity and why slash 24 blocks matter

Two thousand addresses that act like eight

Picture a trial list with two thousand proxy addresses on it. The dashboard says two thousand, the pricing page says the pool is huge, and the first few hundred requests go through without trouble. Then one morning a large share of them start failing on the same target at the same moment, even though no single address was doing much.

When you sort that list, the explanation is often sitting in the first three numbers of each address. Two thousand addresses can fit inside eight /24 blocks, since each block holds 256 of them. To a system that judges traffic by its neighbourhood, that pool never had two thousand members. It had eight.

That’s the whole idea behind subnet diversity, and it’s the part of pool quality that address counts hide. Other pieces on this site deal with how many unique exits a pool really has and why the headline pool size is close to meaningless. Both of those count addresses. This one is about how the addresses are grouped, which is a separate question and often the one that decides how a pool behaves under real load.

There’s a reliability reason to care and a buying reason. The reliability reason is that anything applied to a range, whether a rate limit, a blocklist entry or a stricter rule, lands on every address in that range at once. The buying reason is simpler: if you’re paying for a pool, you should know what you’re getting, and the grouping is part of what you’re getting.

Nothing below is about slipping past a site’s limits. A range-level limit that trips because of your own traffic is the site telling you to slow down, and more subnets won’t change that answer. What subnet spread does protect you from is someone else’s history, and a false picture of what you bought.

What a /24 actually is

An IPv4 address is four numbers between 0 and 255, written with dots. A /24, said “slash twenty-four”, is the block of addresses that share the first three numbers. So 203.0.113.0 through 203.0.113.255 is one /24: 256 addresses, of which a few are usually reserved, leaving about 254 usable hosts.

The notation comes from CIDR, which describes a block by how many leading bits are fixed. Twenty-four fixed bits leave eight free bits, and eight bits give 256 combinations. A /16 fixes only the first two numbers and holds 65,536 addresses. A /22 is four adjacent /24s.

The /24 matters for historical and practical reasons. For a long time it was the smallest block that most of the internet’s routing system would accept as a separately announced route. Networks tend to be allocated, assigned and managed in /24 chunks. A hosting provider rents a customer a /24, or a slice of one. An ISP assigns a neighbourhood’s worth of home connections from a handful of /24s. So the /24 lines up with real ownership and real management more often than any other boundary.

That doesn’t make it a perfect proxy for ownership. One /24 can be carved into smaller pieces and leased to several customers. Several adjacent /24s can belong to one company. But as a default unit for “these addresses probably share a fate”, the /24 is the one most systems reach for, because it’s cheap to compute and right often enough.

IPv6 works differently, and it comes up further down. For now, keep the IPv4 picture: the last number in the address is the house, the first three are the street.

Why targets and reputation systems think in neighbourhoods

A site trying to manage automated traffic faces a practical problem. There are about four billion IPv4 addresses. Tracking reputation and rate limits per address is possible, but abuse rarely stays on one address. An operator with a block of servers rotates through them. A compromised hosting account spins up new machines on nearby addresses. So a defence that only watches individual addresses keeps getting surprised by the address next door.

Grouping by /24 solves that cheaply. If ten addresses in the same /24 misbehave, the eleventh gets less benefit of the doubt. A firewall rule that blocks a whole /24 takes one line instead of 256. A rate limiter keyed on the /24 prefix sees a rotating scraper as one client, which is often exactly what it is.

Public blocklists do this too. Several well-known spam and abuse lists publish entries as ranges, and some escalate from single addresses to a whole range when problems spread across it. Commercial reputation feeds commonly score at more than one level: the address, the /24, the network block, the autonomous system. The piece on rate limiting by ASN covers what happens when a target moves up to that last level.

Most sites don’t publish their exact rules, so I can’t tell you which unit a given target uses. Some key on single addresses. Some on /24s. Some on ASNs or on a mix with other signals. What’s safe to assume is that grouping exists somewhere in the stack of most serious bot management, and that a pool designed as if every address stood alone is designed for a world that mostly doesn’t exist.

For a scraper, the consequence is simple. Whatever happens to one address in a range may happen to the range. A pool’s real size, for this purpose, is closer to its number of distinct neighbourhoods than its number of addresses.

How a crowded pool behaves under load

Take two pools with the same address count. Pool A has 2,000 addresses spread across 1,500 different /24s. Pool B has 2,000 addresses packed into eight /24s. On a quiet target with per-address limits only, you won’t see much difference between them. Both hand out a fresh address when asked.

Now apply range-level treatment. Pool B’s traffic is concentrated: every request from every address lands in one of eight buckets. A target that rate limits per /24 sees eight busy clients. A reputation system that notices one of those ranges misbehaving, perhaps because of another customer on the same pool, can downgrade an eighth of your capacity in one step. A blocklist entry on one range removes 250 addresses at once.

Pool A under the same treatment degrades gradually. One range gets flagged and you lose one or two addresses. The failure is local.

That difference is the thing to understand. Concentration turns small, local problems into large, sudden ones. It also makes your own monitoring harder to read, because a range-level block shows up as a wave of failures across many addresses at once, which looks like a provider outage or a target change rather than a reputation event.

There’s a subtler cost too. Rotation logic that picks the next address in list order, which is common in simple setups, will often step through a range sequentially. So consecutive requests come from 203.0.113.14, then .15, then .16. From the target’s side that looks like one client walking along a street. If your provider’s gateway rotates for you, you can’t see the order at all without logging the exits.

None of this means a concentrated pool is useless. For targets that only care about individual addresses, it’s fine. It just means the address count on the invoice overstates what the pool does against targets that think in neighbourhoods.

Why datacenter pools are usually the most concentrated

The type of pool shapes its subnet spread more than anything else.

Datacenter pools are built from address space a provider leases or owns in blocks. A provider buying a /22 gets four contiguous /24s and fills them with proxy addresses. So datacenter pools tend to be dense: many addresses per range, few ranges, and those ranges all announced by a handful of hosting networks. That’s efficient to run and cheap to sell. It’s also the pattern that range-level systems catch most easily, because the entire range is obviously hosting space and obviously full of proxies.

Some datacenter providers spread their pool across many small allocations from different upstream networks. That costs them more and is worth asking about. A datacenter pool with genuine spread across ranges and networks behaves very differently from one sitting in two /22s at one host.

Static residential offerings, sometimes sold as ISP proxies, are a related case. The addresses are registered to consumer or business ISPs, but they’re often announced from a small number of leased ranges. So they can look residential at the registration level while being as concentrated as a datacenter pool at the subnet level. The only way to know is to look at the list.

The same concentration shows up on the reputation side. If a datacenter range is full of proxy addresses and one tenant abuses a target, the target’s simplest response is to block the range. Every other tenant in it inherits that decision. The shared pool piece on this site calls this the part of sharing that isn’t on anybody’s pricing page, and the subnet is where it physically happens.

Residential pools and the illusion of spread

Rotating residential pools usually do better on raw subnet count, because the exits are real home connections scattered across many ISPs and many ranges. A large pool can easily span thousands of /24s.

But raw count isn’t the same as useful spread, for three reasons.

First, residential ranges are shared with real people. A /24 at a big consumer ISP may hold a couple of hundred households. If a residential proxy network has several exits in that range, and one of them gets used for abuse, the target may downgrade the range. Your request, the network’s other exits, and the ordinary family watching videos on the same range all get worse treatment. So residential spread is real, but the neighbourhood includes people you’ll never see.

Second, availability is lumpy. Residential exits come and go as devices go online and offline. The pool’s spread at three in the afternoon in one country can look nothing like its spread at three in the morning. A test at one moment gives you a snapshot, not a constant.

Third, country and city targeting shrink the pool. A pool spread across thousands of ranges globally may have a few dozen ranges in the city you actually need. Filter your test to the geography the job will use, or the numbers mean nothing for you.

So residential pools often look well spread in aggregate and turn out more concentrated than expected once you narrow to one country and one time of day. That’s worth measuring for the exact slice you plan to use.

Mobile exits break the counting model

Mobile proxies are the odd case, and it helps to understand why before counting anything.

Mobile carriers put huge numbers of subscribers behind a relatively small number of public addresses using carrier-grade NAT. The piece on carrier NAT on this site covers the mechanics. The short version is that one public address may be shared by many phones at once, and the carrier’s public addresses sit in a limited set of ranges assigned to its gateways.

So a mobile pool from one carrier in one country will, by its nature, span few /24s. That’s normal. It’s how the carrier network is built, and every real phone on that carrier lives in the same ranges. Sites know this, which is why they’re cautious about range-level action against mobile carrier space: blocking one carrier range can cut off a large number of genuine customers.

What that means for measurement is that subnet diversity is the wrong yardstick for mobile exits. A mobile pool with ten /24s isn’t “concentrated” in the way a datacenter pool with ten /24s is, because each of those ranges is carrying a crowd of ordinary users as well as your traffic. The right questions for mobile are which carriers, which gateways, how rotation is triggered, and how the provider keeps its own traffic from dominating any one address. The piece on rotating too fast on mobile explains why the last of those matters to the carrier as much as to the target.

Don’t apply datacenter counting rules to a mobile list and conclude it’s poor. Measure what matters for the type.

Measuring subnet spread on a trial list

Here’s the part you can do yourself, during any trial, in about an hour.

First, collect the addresses you actually exit from, not the addresses on the provider’s list. For a rotating gateway, send a few hundred requests through it to an echo endpoint that returns your public address, spaced out and ideally over a day, and log each result. For a static list, send one request through each entry and log what comes back. Lists and exits don’t always match: some entries route through the same upstream, and some rotate behind the scenes.

Second, group the addresses by /24. In Python the standard library’s ipaddress module does this directly: build an IPv4Network from each address with a /24 prefix and strict mode off, then count how many exits fall in each network. Do the same at /16 for a coarser view. If you can map each address to its announcing network through a public lookup, count distinct networks too.

Third, look at three numbers. The count of distinct /24s. The share of all exits that fall in the biggest ten /24s. And an effective count: square each range’s share of the total, add the squares, and divide one by the result. That last number says how many equally sized ranges your pool is worth. A list of 2,000 exits spread evenly over 500 ranges has an effective count of 500. The same 2,000 with half of them in one range has an effective count below four, however many other ranges it touches.

Fourth, repeat for the geography and the time of day your job will use. Global numbers flatter residential pools.

Keep the raw log. A month later, run the same measurement again. A pool whose spread shrinks between measurements is telling you something about where its supply comes from.

Working the arithmetic on a small example

Numbers make this easier to see, so here’s a made-up list worked by hand. These figures are an illustration of the method, not a measurement of any provider.

Say your exit log from a trial holds 400 distinct addresses. You group them by /24 and find 40 ranges. That sounds healthy: ten addresses per range on average. But the distribution matters more than the average. Suppose one range holds 200 of the addresses, three more hold 40 each, and the remaining 36 ranges hold the last 80 between them, a couple each.

Now the effective count. The big range has a share of 200 out of 400, which is 0.5, and 0.5 squared is 0.25. Each of the three mid-sized ranges has a share of 0.1, squared to 0.01, so 0.03 for the three. Each small range has a share of about 0.005, squared to 0.000025, and 36 of those add up to about 0.0009. The total is roughly 0.281. One divided by that is about 3.6.

So a list of 400 addresses across 40 ranges is worth about three and a half equal neighbourhoods. Half your capacity lives on one street. If that range gets listed or rate limited, half your pool goes with it in one step, and the 36 small ranges can’t carry the job alone.

Compare a second trial with 300 addresses across 150 ranges, two per range. Its effective count is 150. Fewer addresses, far more resilience against range-level events.

That’s why the address count is the least useful of the numbers. Put the effective count next to it in any comparison, and look at the share held by the biggest range, and the two trials stop looking alike.

Reading the result without fooling yourself

The numbers from that exercise don’t come with a pass mark, and anyone who gives you a universal threshold is guessing. What they give you is a comparison and a warning.

As a comparison, run the same test on two or three trials and put the numbers side by side. The distinct-range count and the effective count will separate providers faster than any sales conversation. Two pools with the same address count and very different effective counts will behave very differently against range-aware targets.

As a warning, look at the top of the distribution. If a large share of your exits sits in a handful of ranges, one bad event in one of those ranges takes out that share at once. Decide whether your job can survive that. A small daily scrape probably can. A job that feeds a live dashboard probably can’t.

Then weigh the type. A datacenter list with a low effective count is the classic concentrated pool. A mobile list with a low effective count is normal, and you should judge it on carriers and gateways instead. A residential list should show a high count; if it doesn’t in your target country, the country filter has shrunk it more than the pricing page suggested.

Watch for the list that’s padded. A provider can inflate an address count by listing every address in a few ranges, some of which route through the same upstream exit. Your exit log catches this: if 500 list entries produce 60 distinct exit addresses, the list is describing configuration, not capacity.

Finally, remember that spread only tells you about exposure to range-level events. It says nothing about whether a pool’s addresses have clean histories, which the IP reputation piece on this site deals with, or whether the provider sources them ethically, which is a separate question to put to the sales rep.

Rotation and cooldowns that respect the neighbourhood

Once you know the pool’s shape, a few changes to how you use it follow directly.

Track health per range as well as per address. If you already keep a cooldown per address per target, add a second counter per /24 per target. When several addresses in the same range start failing on the same target, treat the range as affected and pause it for that target. That turns a wave of confusing failures into one clear event in your logs.

Avoid sequential rotation through a range. If you control the order, shuffle across ranges so consecutive requests don’t march through adjacent addresses. If the provider’s gateway controls rotation, log exits for a while and see what order it uses.

Spread concurrency across ranges. If you run several workers, give each a different range where possible. The aim is resilience: keep a single range from carrying all of your load, so one range-level event doesn’t stall every worker at once.

And set your own pace per target, independent of pool shape. A well spread pool doesn’t license more traffic to one site. The polite rate for a target is the rate its owner would accept, whatever your subnet count. If a range-level limit trips while you’re inside your own rate budget, you may be sharing the range with someone noisier. If it trips because your total traffic is high, it was aimed at you, and the answer is to reduce the traffic.

When a concentrated pool is the right buy

It would be easy to read all this as “always buy the most spread pool”. That’s wrong for plenty of jobs.

If your targets judge traffic per address, or barely judge it at all, spread buys you nothing you’ll notice. The docs sites, open data portals and small shops that serve plain datacenter traffic without complaint are exactly the targets where a dense datacenter block does the job. Paying more for spread there is the same mistake as paying for residential when you didn’t need it.

Concentration can even help. A small, stable block of addresses you control is easy to allowlist. If a site owner agrees to let your crawler in, or a partner gives you API access tied to source addresses, you want a short, fixed list inside one or two ranges, not ten thousand exits scattered across the internet. The same applies to your own infrastructure and to any job where the other side needs to recognise you.

Stability is another reason. A dense range from one provider tends to have predictable routing and latency. A wide residential spread has more variance in every direction.

So the question to ask is whether your targets think in neighbourhoods and whether your job can survive losing one. If both answers are yes, spread is worth paying for. If the targets don’t care, or the job is small enough to pause for a day, a concentrated pool may be the cheaper and saner choice.

The IPv6 version of the same question

IPv6 changes the arithmetic so much that counting addresses stops making sense at all.

An ordinary IPv6 subscriber, whether a home connection or a single server, is typically given at least a /64, which holds about eighteen quintillion addresses. Many get a /56 or a /48, which hold vastly more. So a single machine can produce an effectively unlimited number of distinct IPv6 source addresses. A provider advertising “millions of IPv6 addresses” may be describing one or two customer allocations.

Sites that handle IPv6 sensibly know this, and they group by prefix: commonly the /64, often the /56 or /48. Treating each IPv6 address as a separate client would make per-address rate limiting pointless, so few serious systems do it.

For measurement, the same method applies with different boundaries. Group your logged IPv6 exits by /64 and by /48, count distinct prefixes, and compute the same effective count. A pool of a million IPv6 addresses in one /48 is, for most purposes, one neighbourhood.

There’s also a practical wrinkle. Some targets don’t serve IPv6 at all, and some treat IPv6 traffic differently from IPv4. A dual-stack pool might send some requests over each depending on client settings. Make sure your exit log records which protocol each request actually used, or you’ll be averaging two very different pools into one misleading number.

Questions to put to a provider

Measurement during a trial beats anything a provider tells you, but a few questions are still worth asking, and the answers tell you something even when they’re vague.

Ask how the pool is spread across ranges and networks, and whether they can share a count of distinct /24s for the country you need. A provider that tracks this will usually have a number. One that answers only with a total address count may not be tracking it.

Ask whether datacenter or static residential addresses come from their own allocations or from leased ranges, and from how many upstream networks. More upstreams usually means more spread and less shared fate.

Ask what happens when a range gets listed. Do they rotate it out, request delisting, or leave it in the pool? A provider that watches for listed ranges and pulls them is doing work that saves you failures.

Ask how they handle one customer abusing a shared range. The answer tells you how much of your fate depends on strangers.

Ask, too, whether the addresses you get during a trial come from the same pool paying customers use. Some providers route trial traffic through a separate, smaller slice. Measuring that slice tells you little about the real product, so it’s fair to ask, and fair to repeat the measurement in the first week of a paid plan.

And ask how often the pool’s ranges change. A pool that keeps adding fresh ranges and retiring old ones can look well spread in one snapshot and different a month later. That’s not necessarily bad, since retiring listed ranges is healthy, but it means your one-off measurement has a shelf life.

Then check the answers against your exit log. The log is the only version of the truth that doesn’t depend on anyone’s incentives, including mine.

What I can’t tell you

I can’t tell you what spread a specific target needs, or which grouping unit it uses. Sites don’t publish that, and it changes. The effective count gives you a way to compare pools, not a threshold that guarantees anything.

I also can’t tell you that a well spread pool will keep you from being blocked. Range-level treatment is one mechanism among many. Request rate, client fingerprint, behaviour and the target’s own rules all sit on top of it, and none of them care how many subnets you have. A job that’s too aggressive gets stopped regardless of spread, and it should be.

What subnet diversity does tell you is how much of your capacity sits behind each single point of failure. That’s worth knowing before you commit money, and it’s easy to measure yourself: log your real exits, group them by /24 or by /64, compute the effective count, and compare trials on the same terms.

If you want the companion pieces on pool uniqueness, pool size claims and how targets treat whole networks, Proxy Scraping has the full set of guides from the operator side.

Get new guides and videos first — join the Telegram channel.

proxies
Need proxies that survive the block wall?

Singapore Mobile Proxy runs real 4G/5G mobile IPs on rotating SIMs — the carrier-grade addresses most of these targets still trust.

see plans →
read on
More scraping guides

The rest of the field manual: target-site playbooks, library walkthroughs, provider reviews, and anti-bot troubleshooting.

browse all guides →