On this page 10 sections
- What actually separates the four types
- What a target sees: four real lookups
- The decision order
- 1. Test whether the target checks network origin at all
- 2. Decide whether you need a stable identity
- 3. If you are blocked, diagnose before upgrading
- 4. Escalate one step, not three
- Where ISP and mobile proxies fit
- Cost behaves differently per type
- Reducing what you transfer
- Running your own comparison
- Signals you are solving the wrong problem
- Questions worth asking before you buy
- How the lookups were run
Start with datacenter proxies and move to residential proxies only when a test shows the target rejects datacenter addresses. The two route your traffic the same way, as any proxy server does. What differs is who owns the exit address: a hosting company, or a consumer internet provider. That ownership is public, and a site can look it up for free on every request. When we ran the same free lookups on an AWS address, a Hetzner address, a Comcast address and a T-Mobile address, both classification databases we checked, ip-api and ipinfo, labelled the first two as hosting and the other two as not. That label is most of what you pay residential prices to change.
What actually separates the four types
Every public address belongs to a network identified by an autonomous system number, and the registry records who holds it. It is cheap for a filtering system to check and hard for a client to fake. The four proxy categories are really four answers to the question “whose network is this exit on?”
| Type | Address registered to | Traffic path | Usual billing | Identity over time |
|---|---|---|---|---|
| Datacenter | Hosting or cloud company | Server to target | Per address | Static |
| ISP (static residential) | Consumer ISP, but hosted in a datacentre | Server to target | Per address | Static |
| Residential | Consumer ISP | Relayed through a real home connection | Per gigabyte | Rotating by default |
| Mobile | Mobile carrier | Relayed through a carrier network, often behind CGNAT | Per gigabyte | Carrier-controlled |
The traffic path explains the practical differences people notice. A datacenter or ISP exit sits on a server network, so the path to the target is server-grade end to end. A residential exit relays your request through someone’s home connection, so it inherits that connection’s upload speed and uptime, and it can disappear when the device goes offline. That is also why residential pools rotate by default: the supply itself comes and goes.
What a target sees: four real lookups
None of the addresses below is a proxy exit. Each one stands in for an owner type, picked from a public source so the result can be reproduced: an address from Amazon’s published EC2 ranges, the server behind hetzner.com, Comcast’s public DNS resolver, and the first address of a prefix announced by T-Mobile USA. We asked three free sources about each on 21 October 2026.
curl -s -X POST "http://ip-api.com/batch?fields=query,isp,as,mobile,proxy,hosting" \
-d '["52.4.0.1","213.133.116.44","75.75.75.75","172.56.8.1"]'
{"isp":"Amazon.com, Inc.","as":"AS14618 Amazon.com, Inc.","mobile":false,"proxy":false,"hosting":true,"query":"52.4.0.1"}
{"isp":"Hetzner Online GmbH","as":"AS24940 Hetzner Online GmbH","mobile":false,"proxy":false,"hosting":true,"query":"213.133.116.44"}
{"isp":"Comcast Cable Communications","as":"AS7922 Comcast Cable Communications, LLC","mobile":false,"proxy":false,"hosting":false,"query":"75.75.75.75"}
{"isp":"T-Mobile USA, Inc.","as":"AS21928 T-Mobile USA, Inc.","mobile":true,"proxy":false,"hosting":false,"query":"172.56.8.1"}
The routing registry gives the same split from a different angle. RIPEstat’s prefix overview reports which network announces each address:
curl -s "https://stat.ripe.net/data/prefix-overview/data.json?resource=172.56.8.1"
# data.resource: 172.56.8.0/23
# data.asns: [{"asn": 21928, "holder": "T-MOBILE-AS21928 - T-Mobile USA, Inc."}]
| Address | Picked from | Announced by (RIPEstat) | ip-api hosting | ip-api mobile | ipinfo Hosting |
|---|---|---|---|---|---|
| 52.4.0.1 | AWS EC2 range 52.4.0.0/14, us-east-1 | AS14618 Amazon | true | false | Yes |
| 213.133.116.44 | A record of hetzner.com | AS24940 Hetzner Online | true | false | Yes |
| 75.75.75.75 | Comcast public DNS resolver | AS7922 Comcast | false | false | No |
| 172.56.8.1 | Prefix 172.56.8.0/23 announced by AS21928 | AS21928 T-Mobile USA | false | true | No |
ipinfo went one step further on the last address and named the carrier with its mobile network code, “T-Mobile MCC 310 MNC 260”. ip-api returned proxy: false for all four, which is correct and also a reminder: its documentation defines that flag as “Proxy, VPN or Tor exit address”, a list of known exits, not a network type. The network type is the separate hosting field, defined as “Hosting, colocated or data center”.
Two things follow. First, these are free, keyless lookups that any site can run on every request, so assume your target knows which column your exit falls in. Second, clean cases are easy. The interesting case is the ISP proxy, where the registry says “consumer ISP” while the hardware sits in a datacentre. Each database decides that case on its own evidence, so check your own exit addresses with the same three lookups before you assume how a target will label them.
The decision order
Run these in order and stop at the first step that works.
1. Test whether the target checks network origin at all
Some targets never look. Two requests tell you whether yours does:
for P in "" "http://user:pass@dc.example.net:8000"; do
curl -s -o /tmp/body -w "%{http_code} %{size_download} bytes ${P:+via proxy}\n" \
${P:+-x "$P"} https://target.example/product/123
done
Compare the byte counts as well as the status. A 200 that returns a fraction of the normal size is a block page, and status alone would score it as a success. If both runs return the full page, use datacenter. It is the cheapest option that works, and nothing further is needed.
2. Decide whether you need a stable identity
Logging in, holding a basket or paging through results needs one address for the duration. That means an ISP proxy, or a residential proxy with a sticky session. Changing address mid-flow is itself a signal, as covered in sticky vs rotating proxies.
3. If you are blocked, diagnose before upgrading
This is where money gets wasted. Before concluding you need a better pool, establish that the address is what failed. If an IP echo service answers through the same proxy while your target refuses you, the proxy is healthy and something about your client is the problem. Check, in order:
- Your TLS ClientHello: does your client handshake like the browser it claims to be?
- Header set and order, compared against what real clients send.
- Whether the page needs JavaScript to issue a token.
- Request rate. A 429 is pacing, not address quality.
Only when all four are clean is the address the likely cause.
4. Escalate one step, not three
Datacenter to ISP can be enough, and it keeps per-address billing and a server-grade path. Go to residential when you need volume across many locations. Go to mobile only when residential has demonstrably failed, or the target is a mobile-only service.
Where ISP and mobile proxies fit
ISP proxies exist to combine the two columns above: a consumer-ISP registration with datacentre hosting. They suit long sessions on one identity, such as managing an account, where a residential exit might vanish mid-task. The risk is that a range gets classified as hosting by the databases a target uses, and then every address in it degrades at once.
Mobile addresses are hard to block for a structural reason: carrier-grade NAT puts many real subscribers behind each shared address, so blocking one address can block many genuine users at once. That is the advantage you pay for. The cost is a carrier radio link in every request’s path, and per-gigabyte billing.
Cost behaves differently per type
Per-address and per-gigabyte plans are not directly comparable, and which is cheaper depends on your traffic shape.
- High bandwidth, few identities, such as rendering pages or downloading files: per-address billing usually wins, because you are not metered per byte.
- Low bandwidth, many identities, such as small API responses from many regions: per-gigabyte residential usually wins, because you pay for very little data.
A headless browser on a per-gigabyte plan is the classic expensive mistake. It fetches every image, font and script by default, and each byte is billable. Measure one real page through the proxy you intend to buy before sizing a plan:
curl -x "$PROXY" -s -o /dev/null --compressed \
-w 'downloaded: %{size_download} bytes total: %{time_total}s\n' \
https://target.example/product/123
Reducing what you transfer
On per-gigabyte billing, every byte you do not download is a byte you do not pay for. Compression is the easiest win, but only if the server honours it. We fetched two public pages three ways with curl 8.7.1 on 21 October 2026, asking for no compression, gzip, and Brotli only:
| Page | No compression | gzip | Brotli only |
|---|---|---|---|
| RFC 9110 (rfc-editor.org) | 1,187,554 | 212,574 | 201,255 |
| Wikipedia: Proxy server | 298,919 | 56,665 | 298,919 |
gzip cut the RFC page to about 18 percent of its size and the Wikipedia page to about 19 percent. The Wikipedia row is the useful surprise: asked for Brotli alone, it sent the full uncompressed page, on both runs. A client that only advertises an encoding the server does not use pays full price without any warning. Send Accept-Encoding: gzip, br, or use curl --compressed, and check size_download rather than assuming.
Other changes worth making, unquantified here because the effect depends entirely on the page:
- Block images, fonts and media when rendering with a browser.
- Use a JSON endpoint instead of the rendered page where one exists.
- Use HEAD for availability checks. For the RFC page, HEAD returned 641 bytes of headers instead of a 1,187,554-byte body.
- Stop re-fetching unchanged content; honour
ETagandLast-Modified.
Running your own comparison
Vendor claims are a starting point, not evidence. During a trial, test each candidate against your real target at the same moment:
#!/usr/bin/env bash
# Tested with bash 3.2.57 and curl 8.7.1.
URL="https://target.example/product/123"
MIN_BYTES=20000 # set this from a known-good fetch of the same page
for P in "http://user:pass@a.example.net:8000" "http://user:pass@b.example.net:8000"; do
ok=0
for i in $(seq 1 25); do
read -r code size < <(curl -s -o /dev/null --max-time 20 -x "$P" \
-w '%{http_code} %{size_download}\n' "$URL")
[ "$code" = 200 ] && [ "$size" -ge "$MIN_BYTES" ] && ok=$((ok+1))
done
printf '%-22s %2d/25 usable\n' "${P##*@}" "$ok"
done
It counts a 200 and a plausible size, so block pages do not score as successes, and ${P##*@} prints the gateway without the password. Run against local test proxies, one of them pointing at a closed port, it printed 3/3 usable for the working ones and 0/3 usable for the dead one, which is the control that proves the script can fail.
Twenty-five requests will not separate close candidates. It is enough to eliminate the obviously unsuitable, which is the decision most people need to make during a trial. Repeat on another day before trusting the order, because pools change.
Signals you are solving the wrong problem
| Signal | What it points to | Next step |
|---|---|---|
| An IP echo works through the proxy; the target refuses you | Your client, not the address | Compare TLS fingerprint and headers with a real browser |
| Connection reset with no HTTP status | TLS-level rejection | See the ClientHello study |
| Mostly 429s | Pacing or an account-level limit | Slow down; more expensive addresses will not help |
| Datacenter to residential made no measurable difference | The address was never the variable | Go back to step 3 |
| 200s that are much smaller than a normal page | Soft blocks or challenge pages | Score by size, not status |
| Results degrade across a whole range at once | The range was reclassified as hosting | Re-run the three lookups on your exits |
Questions worth asking before you buy
- How many addresses are available in the specific countries I need, rather than the global pool figure?
- Is the advertised pool size measured now, or accumulated over a period?
- What is the concurrency limit on the tier I am buying, and what happens when I exceed it?
- Does unused traffic roll over or expire?
- Is session control available on this tier, and what is the maximum window?
- Can I target by ASN, or only by country?
- Can I test against my real target during the trial?
The last one matters more than the rest combined. An IP echo service runs no anti-bot system, so a trial that only works against one tells you nothing about your target. For configuration, see using a proxy with curl and proxies in Python.
How the lookups were run
All measurements on this page were taken on 21 October 2026 from one machine. The AWS address came from ip-ranges.amazonaws.com/ip-ranges.json (createDate 2026-10-21-16-37-05), taking the first address of an EC2 prefix in us-east-1. The T-Mobile address is the first address of a prefix listed by RIPEstat’s announced-prefixes endpoint for AS21928. ip-api was queried once with its batch endpoint, RIPEstat once per address with prefix-overview, and ipinfo once per address from its public HTML page. Each source was queried once, so this shows how four clear cases are labelled, not how accurate any database is.
The compression figures are single downloads per encoding, repeated once. Sizes matched on both runs except the RFC page’s Brotli response, which was 201,255 bytes on the first and 201,399 on the second.
Frequently asked questions
Are residential proxies always better than datacenter proxies?
How can a website tell a datacenter proxy from a residential one?
What is the difference between ISP and residential proxies?
When are mobile proxies worth it?
Why did switching to residential proxies not fix my blocks?
Sources
- IANA: Autonomous System Number registry
- RIPEstat Data API: prefix-overview
- ip-api: JSON endpoint and the hosting, mobile and proxy fields
- ipinfo: IP address data, including the hosting flag
- AWS: published IP address ranges
- RFC 9110: HTTP Semantics (content codings and HEAD)
- RFC 6888: Common requirements for carrier-grade NATs