Proxy DetectionREVIEW / 2026

The published framework

Proxy detection methodology

All ten tools share 42 criteria across seven categories. The framework puts coverage, legitimate-user impact and fresh evidence ahead of a generic proxy flag.

Which tools belong here?

The comparison covers production products that classify anonymized connections or combine network evidence with an observed device and visit. It includes IP intelligence APIs, browser-assisted detection and downloadable data. A free checker is an evaluation route; the production interface and entitlement are assessed separately.

Datacenter, residential, mobile, VPN, Tor and relay coverage are distinct questions. Hosting alone is not a proxy verdict, and a privacy connection alone does not establish abuse. The intended use cases are signup, login, checkout and other product-access decisions.

Adjacent products and alternatives

Geolocation-only products, public IP lists and forwarding-header heuristics cover narrower tasks. Edge bot-management and physical-location compliance products have different collection and deployment requirements. Consider those interfaces when that is the actual requirement.

Scamalytics, Abstract API, IPHub and vpnapi.io overlap parts of IP risk and anonymizer detection without receiving a score here. DataDome and SEON cover adjacent fraud and bot workflows. The ranked ten are a scoped comparison, not a claim that the market has no other suppliers.

Scoring rubric: 42 criteria in 7 categories

Original category points and complete criterion descriptions, retained from the earlier edition.
CategoryRelative pointsCriteria scored
Anonymizer coverage and residential recall9 criteria28datacenter and hosting classification as an explicit field, not inferred from an ASN name; residential proxy recall that separates a proxied home IP from an ordinary ISP customer; mobile and carrier proxy recall; commercial and self-hosted VPN exits; Tor exits and relays, including Apple iCloud Private Relay; SOCKS4 and SOCKS5 detection without relying on forwarding headers; ISP proxies and compromised-host nodes; a typed verdict naming the anonymizer sub-type; and IPv4 plus IPv6 coverage.
False-positive discipline6 criteria16Apple iCloud Private Relay users; carrier-grade NAT and mobile carrier ranges; corporate VPN, university, hotel, and public Wi-Fi egress; legitimate cloud-hosted users, so a hosting IP is not a proxy by default; confidence as a first-class output, so a low-confidence hit is not auto-blocked; and separating anonymization from abuse.
Evidence freshness and provenance5 criteria14evidence provenance, with honeypots, active probing, and observed exits ranked above resold lists; refresh cadence matched to exits that churn in hours; exposed recency such as first-seen and last-seen; reputation that decays after an address leaves a pool; and speed of recognizing newly rotated exits.
Verdict depth and fraud context7 criteria14a graded score instead of a bare proxy:true ; named reasons behind each classification; proxy-network attribution and connection type; ASN, ISP, and location enrichment in one response; fraud context beyond proxy status, such as bot activity, device reuse, and multi-accounting; fit for signup, login, payment, and scraping decisions; and a documented definition of a positive detection.
Integration, latency, and deployment7 criteria12a production API rather than a paste-an-IP checker; a time to verdict that fits signup, login, and checkout; documentation and a predictable JSON schema; client and server SDKs; webhooks, bulk lookup, and feeds; deployment across hosted API, local database, and JS agent; and uptime and rate limits under bursty traffic.
Device and behavioral corroboration3 criteria8client-side corroboration from timezone, locale, the WebRTC-exposed local IP, and latency; network-stack consistency with the claimed operating system; and composability with device identity and behavioral velocity.
Pricing and access5 criteria6transparent published pricing; a free tier that returns the anonymizer fields rather than geolocation only; self-serve signup without a sales call; cost per lookup at production volume; and licensing that covers commercial use.

The seven category weights total 98 relative points. The original rubric reserves the remaining 2 points for evidence-quality tie-breaking. These are relative weights, not seven percentages totaling 100. The published 0–10 editorial scores are retained; this edition does not introduce a new scoring calculation or invent individual criterion subscores.

Coverage carries the largest share because detecting a known datacenter exit is a different task from recognizing a rotating residential or mobile route. False-positive discipline explicitly includes corporate access, carrier NAT, shared Wi-Fi and Apple Private Relay. Every product is assessed under this same scope.

Earlier study record

The September 2026 edition reported an earlier evaluation of 4 million checks over 39 days. The sample covered datacenter, residential and mobile proxies, VPNs and Tor. It reported 99.5 percent agreement for ShieldLabs on labeled sessions. Those runs were completed before the site’s original September 21 publication.

The earlier text also reported 3 percent false negatives on known proxy exits. That figure uses a proxy-exit subset and is not the complement of agreement across all labeled sessions. The figures retain their original historical scope; this documentation update does not establish a new result on current traffic.

This edition checks current product scope, access and sources and expands practical evaluation guidance. Questions about the historical protocol, labels or denominators can be sent through the editorial contact.

How to read a figure on this site

Evidence notation and the meaning of an unavailable value.
MarkingMeaning
Product factLinked to the relevant vendor page or documentation and checked on the stated date.
Editorial scoreA product assessment on the published 0–10 scale; separate from a vendor risk or confidence score.
Earlier study resultA figure from the original study, attached to that sample and period.
Unknown or unavailableA value or result is not established. It does not mean zero, safe traffic or a negative detection.
Monthly equivalentAn annual invoice expressed as a monthly amount; the billing period remains explicit.

A product page establishes documented capability. A public checker reveals one supported output. A controlled evaluation needs labeled proxy routes, legitimate controls, repeated observations and recorded failures. Read the false-positive guide before interpreting a rate.

Editorial policy

The Proxy Detection Review editorial team maintains this publication, with Chris Donnelly as the named author. Readers can follow product statements to the linked primary sources and use the rubric to inspect the scope behind a score.

Zero paid placements is the editorial rule. A payment cannot buy a score, ranking position or Editor’s Choice award. The award identifies the selected fit for device, network and account-risk context; it does not add scoring points.

Vendor links do not by themselves establish an affiliate agreement. Questions about a link or referral parameter can be sent to the commercial-link channel.

Corrections

Send the page URL, exact passage, current primary source, plan and date. Confirmed product corrections are reflected in the relevant page and review date. A factual correction and a request to change an editorial score are considered separately.

Source directory · About the publication

Reviews scored with this method

Summarize this page with