How to Choose a VPN and Proxy Detection API: 9 Questions
You have a shortlist of VPN detection APIs, most landing pages claim high accuracy, and few say how the service will behave on your traffic. The short answer: ask nine questions about categories, reasons, false positives, freshness, IPv6, limits, storage rights, privacy and pricing, then measure the finalists on a labeled sample of your own logins in shadow mode. A vendor’s accuracy figure is a claim. Your sample is evidence.
This guide is for product and security engineers adding IP intelligence to login, signup or payments. It builds on our explainer of how VPN detection works, which covers the signals behind a verdict.
What a Proxy Detection API Can and Cannot Tell You
A VPN, proxy or privacy detection API takes an IP address and returns a classification, often with a score and category flags. It describes the network behind an address, not a person or an intent, and every answer is probabilistic.
Categories differ in difficulty. Cloud ranges, Tor exits and Private Relay addresses come from published files, so vendors can cover them in principle. Residential proxies are harder. Google’s threat researchers describe networks that route traffic through consumer devices on addresses owned by ISPs, and say this creates significant challenges for defenders who try to detect and block the activity. A block on such an address can also hit a real household, which is our inference, not a finding in the post. Be wary of any vendor that promises complete coverage.
How to Choose a VPN Detection API: The Nine Questions
Send these to every vendor in writing and keep the answers next to your test results.
1. Which categories do you detect, and how do you handle residential proxies?
Ask for the category list as it appears in the response: VPN, proxy, Tor, hosting, relay services such as Apple Private Relay and Cloudflare WARP, and residential proxy. Apple says geo IP providers may mark relay traffic with fields such as is_relay or privacy_proxy, so check that relays are labeled separately. Then ask how residential proxy detection works and what evidence supports it.
2. How is the verdict produced, and do you return reasons?
A bare yes or no is hard to act on. Ask whether the response includes the signals that fired, a category, a score with a defined meaning, and a timestamp. Reasons let you route: a Tor exit and a hosting range deserve different handling, and support needs a reason to explain a challenge.
3. How do you handle false positives, and can I request a correction?
Every detector gets some wrong, so ask what happens next. Is there a documented correction path, who reviews requests, and how long until a fix reaches the API? Do not assume a standard process: RFC 5782, the DNS list specification, leaves listing and removal policy to each operator.
4. How fresh is the data, and how often does it update?
IP data decays. AWS stamps its published range file with a publication time, Google’s cloud.json carries a creation time, and Microsoft updates its Azure file weekly. Ask how quickly a new prefix reaches verdicts and how removed prefixes are handled.
5. Do you cover IPv4 and IPv6 equally?
Do not accept “supports IPv6” as an answer. Ask whether categories, freshness and reasons match for both, and what unit a verdict covers. RFC 8981’s default timers allow a device up to three concurrent temporary addresses, and the interface identifier is 64 bits wide (RFC 4291), so a cache entry keyed to one address may never be hit again. Ask whether results apply to the address or the /64.
6. What are the latency, caching rules and rate limits?
Measure p95 latency from your own region, not the vendor’s. Ask whether you may cache responses and for how long, what the limits are per second and per day, and how they are signaled. HTTP 429 is the standard status for rate limiting, and RFC 6585 lets a server add a Retry-After header, so ask whether the API sends both. These answers set your fail mode, covered below.
7. What can I store, and for how long?
Read the license for caching, retaining results, combining them with your own data and building a derived dataset. Terms can limit storage or forbid it, which collides with shadow mode below. Get answers in the contract, not on a sales call.
8. How do you handle privacy, and what should the data processing agreement cover?
Sending the IP address of a visitor in the EU to a vendor can count as processing personal data. The European Commission lists an IP address among examples of personal data, and GDPR Recital 30 names IP addresses among online identifiers. In its 2016 Breyer judgment (case C-582/14), the Court of Justice held that a dynamic IP address registered by an online media provider is personal data for that provider where it has legal means to identify the person using additional data held by the internet service provider. That ruling applied the older Directive 95/46/EC, not the GDPR. This is not legal advice. Ask your counsel or data protection officer how it carries over to the GDPR, and about your lawful basis, notices, and whether the vendor is a processor or a controller.
Where the vendor acts as your processor, ask for a data processing agreement and check it against Article 28(3) of the GDPR, as explained in the EDPB’s Guidelines 07/2020: processing only on your documented instructions, confidentiality, security measures, sub processor conditions, help with data subject requests, help with breach and impact assessment duties, deletion or return at the end, and information and audit rights. The Commission’s standard contractual clauses for controllers and processors are a ready template, but they state that they do not by themselves ensure compliance with the rules on international transfers. Transfers outside the European Economic Area fall under separate rules in Chapter V of the GDPR. Also ask where queries are processed, how long the IPs you send are kept, and whether they feed the vendor’s own datasets.
9. How is it priced, and what happens when I exceed limits?
Find out what counts as a billable call (retries, errors), whether a free tier exists, and what happens at the cap: errors, throttling, automatic upgrade or overage billing. Model it with real volume, including retries. A cap that returns errors drops you into your fail mode, and overage billing surprises finance.
Fail Open or Fail Closed? Using a VPN Detection API for Login Risk
Two decisions hide in this phrase. First, what happens when the lookup cannot complete because of a timeout, error or quota: fail open lets the request through, fail closed stops it. Second, what you do with a positive verdict: block, challenge or only record. Decide both for each flow.
Four inputs drive the choice: the loss if a bad actor gets through, the cost of a wrong block, how reversible the action is, and whether other signals corroborate. Illustrative numbers show why blocking on IP alone often loses. Suppose 1 in 50 flagged logins is fraud, a fraud costs 200 dollars, and a wrongly blocked customer costs 20 dollars in support and churn. Of 100 flagged logins, 2 are fraud and 98 are not. Blocking avoids 400 dollars of fraud and costs 1,960. A step up check at an assumed 1 dollar per challenge costs 100 dollars in total and still puts a barrier in front of the 2 frauds.
| Flow | If the lookup fails | If the verdict is positive | Why |
|---|---|---|---|
| Login | Fail open, use your other signals | Step up with a second factor or verification, do not block | A lockout hits a real customer, and password and device history still count |
| Signup | Fail open, review later | Challenge, or limit new account privileges at first | New accounts have no history, and limits are reversible |
| Payment | Fail closed for high value, queue for review | Hold for manual review | A bad payment is costly and hard to undo |
How to Test Vendors on Your Own Traffic
Vendor accuracy numbers come from the vendor’s data. Build your own evidence in six steps.
- Build a labeled sample. Pull a few thousand recent sessions and label them by outcome, not by what any detector said: confirmed fraud, chargebacks, takeovers, support verified customers, long standing payers. Fraud is rare, so oversample it or you cannot measure recall. If you run VPN servers, add your own egress IPs as known VPN examples.
- Add free ground truth. Three public sources need no judgment: the Tor Project’s exit list, Apple’s Private Relay egress file and cloud range files such as AWS’s
ip-ranges.json. Use today’s copies. A miss tells you about freshness, but a pass proves little about hard cases such as small VPNs and residential proxies. - Run shadow mode. Call every finalist for each sampled IP and for live traffic for two to four weeks. Log verdict, score, reasons, latency and errors, and take no action on them. Check that question 7’s license allows this, and settle question 8’s privacy terms with your counsel or data protection officer before you send real visitor IP addresses to any vendor.
- Measure precision and recall. Precision is the share of flagged sessions that were truly bad. Recall is the share of bad sessions that were flagged. Also track how many good users get flagged. These numbers are made up to show the arithmetic, not results from any vendor. Your sample has 100 bad sessions and 1,900 good ones. Vendor A flags 80 bad and 40 good: recall 80%, precision 67%, 2.1% of good users flagged. Vendor B flags 60 bad and 10 good: recall 60%, precision 86%, 0.5% flagged. B flags four times fewer good users, and A catches more bad ones. Your fail mode table decides which trade you want. Precision falls as fraud gets rarer, so if you oversampled fraud, reweight to your real rate: at 1% fraud, these two vendors would show about 28% and 54%.
- Review disagreements. Rows where vendors disagree with each other or with your label teach the most. Review about 50 by hand, decide who was right, and keep the notes.
- Rerun quarterly. Ranges, vendors and your traffic change. Keep the sample, add fresh rows, and rerun every quarter and before renewal.
Scoring Sheet
Set weights before you see results. Add a column per vendor, score each row from 1 to 5 and multiply by the weight. Treat privacy terms and storage rights as gates, not points: a vendor that will not sign a suitable agreement or allow the storage you need is out.
| Criterion | Weight | How to measure |
|---|---|---|
| Precision on your sample | 20 | Step 4 |
| Recall on your sample | 15 | Step 4 |
| Good users flagged | 10 | Step 4, labeled good rows |
| Reasons returned | 10 | Question 2 |
| Correction path | 10 | Question 3, submit a real test case |
| Freshness against public files | 10 | Step 2, compare newly published prefixes |
| IPv6 handling | 5 | Question 5, run IPv6 sessions separately |
| Latency and limits | 10 | p95 from your region, Question 6 |
| Price at your volume | 10 | Question 9, modeled on real traffic |
Integration Patterns That Limit the Cost of a Wrong Verdict
Cache by what the verdict covers. Use short TTLs of hours for payments and longer ones for low stakes, and key IPv6 on the /64 if the vendor scores it that way.
Step up instead of blocking. Ask for a second factor or email verification, or slow the action. Apple suggests treating Private Relay addresses like carrier grade NAT and offers Private Access Tokens as a way to protect servers without friction for legitimate users on supported devices.
Keep allow lists for partners. Payment webhooks, corporate gateways, partners and your own staff should skip the check, with an owner and an expiry date on every entry.
Store the reasons and vendor version with each decision, so support can explain it and your next test has real outcomes. Put the vendor behind a feature flag so you can switch to log only during an incident.
FAQ
What is the best VPN detection API?
There is no best one in general. The best is the vendor that scores highest on your own labeled sample at your volume, with terms you can accept. The nine questions and the scoring sheet let you compare finalists.
Can a proxy detection API catch residential proxies?
Partly, and no vendor can promise complete coverage. Residential proxy traffic comes from ordinary consumer ISP addresses, so it is hard to tell apart. Ask for the evidence behind the category and test it on your own confirmed cases.
Is an IP address personal data under GDPR?
It can be. The European Commission lists an IP address among its examples of personal data, and in Breyer the Court of Justice held that a dynamic IP address can be personal data for a website operator with legal means to identify the person, a ruling under the earlier Directive 95/46/EC. This is not legal advice. Your facts decide, so check with your counsel or data protection officer.
Should I block VPN users at login?
Usually no. Many VPN users are legitimate, and a wrong block can cost more than the fraud avoided, as the worked example shows. Step up verification for risky flows and block only where other signals agree.
Takeaway
Choosing a VPN detection API comes down to evidence. Ask the nine questions, decide your fail mode per flow, test on your own traffic, and rerun the test every quarter.
IPLogs is our own product, so weigh this mention accordingly. It is a free VPN, proxy and Tor IP checker with a public REST API, no signup to start, and a verdict plus a 0 to 100 threat score from a 7 layer pipeline across 13 sources, per its product page. We make no accuracy claim here. Put it through the same sample as every other finalist and judge it by your own numbers.
Building IP checks into your VPN app’s signup or billing flow, or not sure where to start? Tell us your market and your traffic on the contact page, and we’ll tell you what we’d do in your position.
Last checked against the Breyer judgment, GDPR Recital 30, European Commission and EDPB guidance, Google’s residential proxy report, Apple’s Private Relay guidance, the Tor Project’s exit list announcement, AWS, Google Cloud and Azure range documentation, and RFC 4291, 5782, 6585 and 8981 on 30 October 2026; the GDPR, Breyer and contract-clause sources were re-checked on 10 November 2026.