Device Fingerprinting vs IP Intelligence: Why You Need Both
Device fingerprinting recognizes a returning browser; IP intelligence reads its network. In our lookups a Mullvad VPN scored 100 and a Private Relay IP 9.

Fraud and trust teams keep staging the same debate: device fingerprinting or IP intelligence, which one should we buy? It is the wrong question. On one large site's real traffic only about a third of fingerprints were unique, they drift within days, and anti-detect browsers spoof them wholesale. The IP arrives with the first packet, server-side, on every request including API and bot traffic, and it condenses into a 0 to 100 IP fraud score. Below: what each layer sees, where each fails, and how to combine them.

What a device fingerprint actually is
A device fingerprint is a hash of attributes the browser is willing to expose. A script reads the user agent, screen, installed fonts, language, timezone and hardware hints, plus the subtle ways a specific GPU and driver render a hidden canvas, WebGL scene or audio buffer. Hash them together and you get an identifier that survives cookie clearing and private windows, because it is derived from the machine rather than stored on it.
That gives fingerprinting three genuine strengths:
- Recognizing a returning browser across sessions, logins and cleared cookies.
- Linking accounts to a shared device, which is a core key in multi-accounting detection.
- Spotting internal inconsistencies that betray automation: a user agent claiming an iPhone with a desktop GPU string, a headless browser with impossible screen dimensions, or a fingerprint that flips on every request.
Used for those jobs, it is a strong signal. The trouble starts when teams treat the fingerprint as an identity rather than a probability.
Three honest limits of fingerprinting
Fingerprinting fails in three ways: it collides at scale, it drifts over time, and it is under attack from fraud tooling, browsers and regulators.
It is far less unique at scale than the famous studies implied. Early research on privacy-conscious volunteers reported over 80% of browsers as uniquely identifiable. When researchers analyzed 2,067,942 fingerprints from one of the top 15 French websites, meaning ordinary traffic rather than people who sought out a privacy test, only 33.6% were unique, and on mobile the figure fell to 18.5% (Hiding in the Crowd, WWW 2018). Millions of people run the same phone model with the same default fonts, so on real traffic most fingerprints collide.
It drifts. Browsers update themselves, fonts get installed, screens get plugged in. The FP-STALKER study, built on 98,598 fingerprints from 1,905 browser instances, found that fingerprints change frequently, "from every few hours to days" (FP-STALKER, IEEE S&P 2018). With its linking method the authors still followed a browser for 54.48 days on average, and 26% for more than 100 days. That is impressive research and also a warning: a naive hash-and-compare system sees the same customer as a stranger within days.
It is under attack from three directions. Fraud rings spoof it deliberately: anti-detect browsers, the standard kit for account farms, generate a fresh and internally consistent fingerprint per profile. Browsers degrade it on purpose: Safari limits what known fingerprinting scripts can read from APIs such as canvas and web audio, and Firefox ships fingerprinting protection that blurs many of the same signals. And regulators treat it as tracking. The EU's Article 29 Working Party concluded in Opinion 9/2014 that the ePrivacy consent requirement applies to device fingerprinting. On 19 December 2024 the UK ICO called Google's move to allow fingerprinting for advertising "irresponsible". Security and fraud prevention have a stronger legal footing than advertising, but the surface is real.
None of this makes fingerprinting useless. It makes it one probabilistic input that needs a second, independent layer to confirm or contradict it.
What IP intelligence sees that a fingerprint cannot
IP intelligence describes the network a request came through, and it does so before any JavaScript runs. The address arrives with the first packet of every request, including API calls, webhooks and headless scripts that never run your fingerprinting code. Enriched, it answers questions the browser can never answer about itself (the field-by-field picture is in what IP intelligence is):
- What kind of network is this?
asn.connection_typeseparates residential and mobile access from datacenter or hosting. A consumer checkout from a cloud instance is a script until proven otherwise, and no fingerprint can tell you that. - Is the connection anonymized?
detection.is_vpn,is_proxy,is_torandis_relay, plus ananonymity_level. The mechanics are covered in how VPN and proxy detection works. - Who operates the network? The ASN and organization, which is how "forty new accounts on one hosting provider" becomes visible.
- Where is it, and does that agree with everything else? Country and city with a
geo_confidence, plus an IANAtimezonewhen one is known. - Does this address have a history? Spam, scanning and botnet flags, and DNS blocklist listings in
threat.dnsbl_sources. - Is it shared?
threat.is_cgnatmarks carrier-grade NAT, where one address is really thousands of people, and the score for those addresses is capped at 50.
Two properties matter most. IP intelligence is server-side and universal: it covers the traffic fingerprinting is blind to. And it is hard to fake from the client: a request has to come from somewhere, and a datacenter or a known VPN server shows up as one. One caveat sits on your side: score the real client IP from your trusted proxy chain, never a spoofable forwarded header.
What IP intelligence cannot do alone
An IP address cannot tell two customers apart when they sit behind the same CGNAT address, office gateway or campus network. It cannot recognize the same laptop when it moves from home Wi-Fi to a phone hotspot, and mobile addresses rotate constantly.
The hardest gap is the residential proxy. A fraud ring that rents exits on real home connections sends traffic that looks like an ordinary home user, and IP data mostly cannot see it. We do not detect residential proxies, and no IP lookup should be your only defense against them. This is exactly where the fingerprint earns its place: the same anti-detect profile reappearing across many "home" addresses is a device-layer pattern, not a network one. So IP intelligence alone cannot answer "is this the device we trusted last week?", which is the question fingerprinting exists to answer.
Question by question: which layer answers it
Each layer answers the questions the other cannot, and the two lists below make the split explicit.
Fingerprint
- Have I seen this browser before?: Yes (probabilistic)
- Are these accounts on one device?: Yes
- Is this a datacenter, VPN, Tor or relay connection?: No
- Is this a residential proxy exit?: Indirectly (same device, many homes)
- Which network operator, and what history?: No
- Where is the user, roughly?: Weakly (timezone, language)
- Does it work on API and bot traffic with no JavaScript?: No
- Does it work before the page loads?: No
- Is it stable over weeks?: No (changes within days)
- Can the client spoof it?: Yes (anti-detect browsers)
- Regulatory exposure: Consent rules apply
IP intelligence
- Have I seen this browser before?: No
- Are these accounts on one device?: Partly (same network)
- Is this a datacenter, VPN, Tor or relay connection?: Yes
- Is this a residential proxy exit?: Mostly no
- Which network operator, and what history?: Yes
- Where is the user, roughly?: Yes (with
geo_confidence) - Does it work on API and bot traffic with no JavaScript?: Yes
- Does it work before the page loads?: Yes
- Is it stable over weeks?: For fixed lines yes, mobile no
- Can the client spoof it?: Only by renting a different exit
- Regulatory exposure: Standard IP processing
Read the two lists together: for almost every question where one layer says "no", the other says "yes".
Same trusted fingerprint, two networks: what we saw on 20 October 2026
The network layer turns one trusted fingerprint into either a block or an allow, depending on where it arrives from. Picture a browser you have trusted for months arriving once from each address below. We looked up the iCloud Private Relay egress 104.28.46.99 at 04:34 UTC and the Mullvad server 185.213.155.73 at 04:35 UTC on 20 October 2026, with no API key. Results change as ranges move.
The Mullvad server in Frankfurt came back like this, trimmed with values untouched:
{
"ip": "185.213.155.73",
"asn": {
"asn": 39351,
"org": "ESAB-AS - 31173 Services AB",
"asn_type": "hosting",
"connection_type": "datacenter"
},
"geo": {
"country_code": "DE",
"city": "Frankfurt am Main (Innenstadt I)",
"timezone": "Europe/Berlin",
"geo_confidence": 0.8
},
"detection": {
"is_vpn": true,
"is_hosting": true,
"is_relay": false,
"is_anonymous": true,
"anonymity_level": "medium",
"vpn_provider": "Mullvad",
"vpn_confidence": 95
},
"scoring": {
"fraud_score": 100,
"recommended_action": "block",
"detection_methods": ["known_vpn_server", "wireguard_open", "datacenter_with_vpn_proxy", "no_ptr_datacenter"]
}
}
The relay egress on Cloudflare's network came back like this:
{
"ip": "104.28.46.99",
"asn": {
"asn": 13335,
"org": "CLOUDFLARENET - Cloudflare, Inc.",
"asn_type": "unknown",
"user_type": "consumer_privacy_network"
},
"geo": {
"country_code": "AO",
"city": "Luanda",
"timezone": "Africa/Luanda",
"geo_confidence": 0.8
},
"detection": {
"is_vpn": false,
"is_proxy": false,
"is_relay": true,
"relay_provider": "iCloud Private Relay",
"anonymity_level": "medium"
},
"scoring": {
"fraud_score": 9,
"recommended_action": "allow",
"detection_methods": ["relay_ip"]
}
}
On 20 October the Mullvad address scored 100, the maximum, and detection_methods says why. Since our 10 November 2026 scoring change a VPN with no abuse, proxy or Tor evidence stops at 70: the same four reasons scored 70, step_up, on 11 November. It is on Mullvad's own published server list, which is where vpn_provider: "Mullvad" and vpn_confidence: 95 come from, and that list tags it as a WireGuard server. A trusted fingerprint arriving from there is the "known device on an anonymized network" contradiction, so step it up even if the device history is clean. The Private Relay address scores 9 with relay_ip as the only method. Apple's relay is a consumer privacy feature, not a proxy, and the score caps relay addresses at 60, so a Private Relay IP can never reach block. A long-trusted fingerprint behind it is about as safe as traffic gets.
Two more lookups from the same run complete the picture:
104.28.46.99 (04:34)
- What it is: iCloud Private Relay egress
fraud_score: 9recommended_action:allowdetection_methods:relay_ip
45.33.32.156 (04:34)
- What it is: Plain Linode cloud server, Fremont
fraud_score: 10recommended_action:allowdetection_methods:datacenter_ip
185.213.155.73 (04:35)
- What it is: Mullvad VPN server, Frankfurt
fraud_score: 100recommended_action:blockdetection_methods: 4 methods, above
185.220.101.1 (04:34)
- What it is: Tor exit, Hamburg
fraud_score: 100recommended_action:blockdetection_methods:tor_exit,datacenter_blocklisted,no_ptr_datacenter,known_spammer
A plain cloud server scoring 10 is a useful reminder: hosting alone is not fraud. A trusted fingerprint arriving from a cloud box deserves a question, not a block.
Two details matter for the timezone check. The Mullvad server reports Europe/Berlin, so a browser clock set to New York on that address is a clean contradiction. The Linode server's geo.timezone came back "none", so on that address the timezone comparison has nothing to compare against. And the relay egress geolocates to Luanda, Angola, while Apple's own egress list put that range in Lusaka, Zambia when we checked on 12 October. A browser clock that disagrees with a relay location is expected and should not count as a proxy tell.
Network verdicts expire, so store the lookup time
An IP verdict is a reading taken at one moment, and our own sample proved it. The earlier version of this post used 102.129.235.17 as its VPN example. On 5 October 2026, and again on 16 October, it returned fraud_score 92, block, with known_vpn_server, openvpn_open and ipsec_open. On 20 October at 04:34 and again at 04:35 UTC it returned fraud_score 0, allow, is_vpn: false and an empty detection_methods. We have not established whether the address left a VPN network or our scan stopped seeing it.
Either way, store the lookup time, the score and the methods next to the fingerprint ID, and re-score the IP on the next sensitive action rather than trusting last month's verdict. A fingerprint drifts within days, and this IP's verdict flipped from block to allow within four days.
Better together: how to combine them
Running both layers pays off where they disagree, and six rules turn that into decisions.
Hunt for contradictions. A fingerprint you have trusted for months suddenly arriving through a datacenter VPN on another continent is a session hijack, account sharing, or a proxy in front of a bot. A brand-new fingerprint on the same residential IP a trusted device used yesterday is almost always a new phone in the household. Same fingerprint plus impossible travel between two IPs in an hour is a stolen session or a rented exit.
Compare the browser's clock with the network's. The fingerprint reports the browser timezone; IP intelligence reports the network's. New York time on a Europe/Berlin IP is one of the oldest proxy tells, and it costs one string comparison. Skip the check when timezone is "none" or the address is a relay.
Use the network to break fingerprint collisions. Two identical fingerprints (same phone, same OS, same defaults) are two different people when their IPs are on different residential networks in different cities. A "changed" fingerprint that keeps arriving from the same IPv6 /64 and ASN is likely the same person after a browser update.
Let each layer cover the other's blind traffic. Fingerprint browser sessions where JavaScript runs, and score the IP on everything.
Cluster on both, independently. For multi-accounting and promo abuse, a shared fingerprint and a shared network are two independent families of evidence. One shared key is a household or a CGNAT block; two independent shared keys are a ring hypothesis worth reviewing.
Score, then step up. Feed the fingerprint in as continuity ("known device" lowers risk, "known bad device" raises it) and the IP as context. Then act in the four bands the score already uses, as in per-action fraud score thresholds: allow at 25 or below, review up to 50, step_up up to 75, and block above 75. When a contradiction lands in review and an analyst has to look by hand, follow the per-incident procedure in how to investigate an IP address.
An implementation checklist
Six build steps turn those rules into a working join.
- Capture the fingerprint on login, signup, checkout and payout, and log its ID with the request.
- Look up the real client IP server-side and keep the
detectionflags, ASN,geo.timezone,is_cgnat,fraud_scoreanddetection_methods. - Store both against the session and account with a timestamp, so you can compare today's pair with last month's.
- Write the contradiction rules first: known device on a VPN, Tor or datacenter network, timezone mismatch, impossible travel, new device on a trusted network.
- Dampen on CGNAT and mobile carrier addresses. A changed fingerprint on a familiar network is not a red flag by itself.
- Document why you fingerprint and confirm the consent position for the regions you serve.
The IP half from a single lookup
Everything the IP half needs comes back from one GET request, and the explainable 0 to 100 fraud score ships with recommended_action and the methods that fired. It also applies exonerating signals: a residential ISP takes 17 points off and a verified crawler 28.5. Every field is on the free tier at 1,500 lookups per day, and you can look up any IP without an API key before you write the join against your fingerprint table.
Stop asking which layer to buy
The pairing shows up across this cluster, from multi-accounting rings to coupon abuse, and it is how you keep detecting anonymized traffic without blocking real customers, because a trusted fingerprint behind Private Relay at 9 deserves a lighter touch than a fresh one on a VPN at 100. The reasons behind each score show which network signals moved it. So don't pick a layer. Ask which questions you need answered, and you will find you need both.
Frequently Asked Questions
What is device fingerprinting?
Device fingerprinting (often browser fingerprinting) collects dozens of attributes a browser exposes, such as user agent, screen size, installed fonts, timezone, language, hardware details, and how the device renders canvas, WebGL, and audio, and hashes them into an identifier. Because it needs no cookie, it can recognize a returning browser across sessions and link several accounts to one device. It runs client-side, so it only sees traffic where its script executes.
Is device fingerprinting accurate?
Less than early studies suggested. A 2018 analysis of 2,067,942 fingerprints from one of the top 15 French websites found only 33.6% were unique, and 18.5% on mobile, against over 80% in earlier studies of privacy-conscious volunteers. Fingerprints also drift: the FP-STALKER study found they change from every few hours to days, though its linking method still followed browsers for 54.48 days on average. Treat a fingerprint as a probabilistic recognition signal, not a stable identity.
Can IP intelligence replace device fingerprinting?
No, and the reverse is also true. IP intelligence sees the network: connection type, hosting or residential origin, VPN, proxy, Tor and relay use, ASN, geolocation, and abuse history. It works server-side on the very first request and on API traffic where no script runs. But it cannot tell two households behind one carrier-grade NAT address apart, and traffic through a residential proxy looks like an ordinary home user. Fingerprinting covers those gaps, and IP intelligence covers fingerprinting's blind spots.
What is the best way to combine device fingerprint and IP signals?
Treat the fingerprint as a continuity signal (have we seen this browser?) and the IP as a context and risk signal (what network is it on, and is it anonymized?), then look for contradictions. A known fingerprint arriving from a datacenter VPN on another continent is a hijack or proxy pattern; a new fingerprint on the same trusted residential IP is usually a new device in the household. Compare the browser's reported timezone with the IP's timezone, and use both as independent clustering keys when hunting multi-accounting.
Is device fingerprinting legal?
In the EU and UK it is regulated as tracking. The Article 29 Working Party concluded in Opinion 9/2014 that the ePrivacy Directive's consent rule for accessing information on a user's device applies to fingerprinting, with narrow exemptions for what is strictly necessary to provide a service the user requested. Fraud prevention has a stronger footing than advertising, but the compliance surface is real: on 19 December 2024 the UK ICO called Google's decision to permit fingerprinting for advertising 'irresponsible'. Fingerprint deliberately, document the purpose, and get legal advice for your jurisdiction.
The team behind the GeoIPHub API
Written by the GeoIPHub team, the people who build the GeoIPHub IP intelligence API and its scanner. Drafts are prepared with AI assistance and edited by the team; API examples are real lookups against the live API unless labelled illustrative.


