Risk Score Reasons: How to Read detection_methods Field by Field
Risk score reasons, field by field: every detection_methods value our scoring code returns, its weight, and a real lookup where a verified crawler cut 64 to 35.

A customer gets blocked at checkout and writes in asking why. "Our vendor said 92" ends the conversation badly. "Your connection came from a listed VPN server with an open OpenVPN port" is an answer you can stand behind, change your mind about, or tune. That is what an explainable fraud score buys: reason codes, not just a number.
This page is the dictionary for those answers. For what the number itself means and where the bands sit, start with what an IP fraud score is and how to read it. Below we show the arithmetic, trace one real response reason by reason, cover the five reasons that subtract, and list every detection_methods value our scoring code can return, with its weight and what makes it fire.

Risk score reasons add up: points times confidence, then a clamp
In GeoIPHub's scoring, every reason contributes its points multiplied by its confidence, and the score is the rounded total, clamped to 0 to 100. There is no hidden model on top. This is the scoring step in our API (scoring.service.ts), reduced to its arithmetic:
// Each fired reason: { points, confidence } from scoring.json.
// scale is 1 except for aging proxy sightings, VPN protocols on a VPN
// and reputation lists on a relay (see below).
let sum = fired.reduce((s, r) => s + r.points * r.confidence * r.scale, 0);
if (isCgnat && !ownThreatEvidence) sum = Math.min(sum, 50); // CGNAT cap
if (isRelay) sum = Math.min(sum, 60); // relay cap
if (isConsumerVpn && !abuseProxyOrTor) sum = Math.min(sum, 70); // VPN cap
if (isProxy && (verified || observedOnly)) sum = Math.max(sum, 26); // proxy floor
const fraud_score = Math.max(0, Math.min(100, Math.round(sum)));
const confidence = unobserved ? 0.05 : fired.length === 0 ? 0.1
: Math.min(0.99, Math.abs(sum) / fired.length / 50 + fired.length * 0.05);
Two details trip people up. First, confidence in the response is not the probability of fraud. It grows with the number of reasons and their average size, so four modest reasons can carry more confidence than one large one. A lookup with no reasons returns 0.1, or 0.05 with scoring.coverage set to unobserved when we hold no evidence on the address at all, so that 0 means no data, not clean. Second, the order of detection_methods is the order the code checks them, not their size. Don't treat the first entry as the main reason.
Other vendors expose reasons differently. MaxMind's minFraud responses carry risk_score_reasons: multipliers on the risk score, each with codes such as ANONYMOUS_IP, and MaxMind notes that they "are usually only returned for medium to high risk transactions." Socure's Sigma Identity Fraud pairs a score between 0.001 and 0.999 with reason codes that "provide detailed information to help you understand the factors that influenced the fraud risk score." GeoIPHub's reasons are additive points, and every reason that moved the score is listed, whatever the score.
We checked that the arithmetic holds on live data. These seven keyless lookups ran on 2026-10-29 between 04:43:09 and 04:43:54 UTC, about seven seconds apart, and every score and confidence rebuilds exactly from its reasons. Re-run on 2026-11-09 between 09:05 and 09:07 UTC after our scoring changed, all seven returned the same reasons, scores and confidence:
185.220.101.1 (Tor exit)
- detection_methods: tor_exit, datacenter_blocklisted, no_ptr_datacenter, known_spammer
- Sum: 108.05
- Score, action: 100, block
- confidence: 0.74
185.213.155.73 (Mullvad, on its list)
- detection_methods: known_vpn_server, wireguard_open, datacenter_with_vpn_proxy, no_ptr_datacenter
- Sum: 101.5
- Score, action: 100, block
- confidence: 0.71
5.39.1.230 (AhrefsBot)
- detection_methods: datacenter_blocklisted, no_ptr_datacenter, verified_crawler, known_spammer
- Sum: 35.0
- Score, action: 35, review
- confidence: 0.375
45.33.32.156 (scanme.nmap.org, Linode)
- detection_methods: datacenter_ip
- Sum: 10.0
- Score, action: 10, allow
- confidence: 0.25
104.28.46.99 (iCloud Private Relay)
- detection_methods: relay_ip
- Sum: 8.5
- Score, action: 9, allow
- confidence: 0.22
107.178.192.1 (Google user-triggered fetcher)
- detection_methods: datacenter_ip, rdns_hosting, verified_crawler
- Sum: -12.5
- Score, action: 0, allow
- confidence: 0.23
66.249.66.1 (Googlebot)
- detection_methods: datacenter_ip, verified_crawler
- Sum: -18.5
- Score, action: 0, allow
- confidence: 0.285
One of the seven has changed since. On 11 November 2026 at 06:19 UTC, after our VPN scoring change went live on 10 November, the Mullvad server returned the same four reasons but scored 70, step_up, with confidence 0.55. Its sum is now 83.1, because wireguard_open counts a quarter on a VPN (6.125), and the VPN cap holds it at 70. The Tor exit, scanme.nmap.org and the iCloud Private Relay egress returned the same reasons, scores and confidence as above.
Tracing one real response: AhrefsBot at 35, review
AhrefsBot's 5.39.1.230 scored 35, review, because three reasons added 63.5 points and its verified crawler match took away 28.5. It is the best teaching case because its reasons pull in both directions. Here is the trimmed response from the 04:43:09 UTC lookup (meta.last_classified_at 2026-10-29T04:43:09.547Z), values unedited:
curl https://api.geoiphub.com/v1/lookup/5.39.1.230 \
-H "x-api-key: YOUR_API_KEY"
{
"ip": "5.39.1.230",
"asn": { "asn": 16276, "org": "OVH - OVH SAS", "asn_type": "hosting" },
"detection": { "is_vpn": false, "is_proxy": false, "is_hosting": true },
"threat": {
"is_spammer": true,
"dnsbl_sources": ["zen.spamhaus.org"],
"crawler": {
"verified": true,
"name": "AhrefsBot",
"operator": "Ahrefs",
"verification_method": "ip_range_list"
}
},
"dns": { "ptr_record": "none" },
"scoring": {
"fraud_score": 35,
"confidence": 0.375,
"recommended_action": "review",
"detection_methods": ["datacenter_blocklisted", "no_ptr_datacenter", "verified_crawler", "known_spammer"]
}
}
Each reason maps to one line of arithmetic:
datacenter_blocklisted
- Why it fired here: OVH hosting address that is also on a spam list
- Points: 30
- Confidence: 0.75
- Adds: +22.5
no_ptr_datacenter
- Why it fired here: Hosting address with no reverse DNS in our records
- Points: 10
- Confidence: 0.50
- Adds: +5.0
verified_crawler
- Why it fired here: Inside Ahrefs' published crawler IP ranges
- Points: -30
- Confidence: 0.95
- Adds: -28.5
known_spammer
- Why it fired here: Listed on zen.spamhaus.org
- Points: 40
- Confidence: 0.90
- Adds: +36.0
Total
- Adds: 35.0, review
Take away the crawler match and the sum is 63.5, which rounds to 64: step up. The verified range file moved a real, declared crawler from a challenge to a logged review. It did not move it to allow, because an exonerating reason is a weight, not an allowlist. A Spamhaus listing on the same address still counts.
Where this response is wrong
One of the four reasons is our gap. no_ptr_datacenter says the address has no reverse DNS, but dig -x 5.39.1.230 at 04:44 UTC returned proxy-fr005-cip7.ahrefs.net. The reason reads our stored PTR records, and that record is missing.
Fixing the gap would not lower this score, and the code shows why. Our PTR classifier reads any hostname containing "proxy" as a proxy pattern. So with the record stored, no_ptr_datacenter (+5) would be replaced by rdns_proxy (+14), and the score would rise to at least 44. That is still review, but for a hostname Ahrefs chose for its crawler egress, not for an anonymizer. We have logged both. Read no_ptr_datacenter and the rdns_* reasons as weak evidence, which is what their weights already say.
Exonerating reasons: five can subtract, and they stop at zero
Exonerating reasons carry negative points, so they lower the sum, but the score never drops below 0. GeoIPHub's scoring config has six of them, and five can fire today:
verified_crawler
- Points: -30
- Confidence: 0.95
- Subtracts: 28.5
- Fires when: Address is inside one of 19 official crawler IP-range files (Googlebot, Bingbot, GPTBot, ClaudeBot, AhrefsBot and others)
residential_isp
- Points: -20
- Confidence: 0.85
- Subtracts: 17
- Fires when: Home broadband line with no VPN, proxy or Tor evidence
rdns_residential
- Points: -20
- Confidence: 0.80
- Subtracts: 16
- Fires when: Reverse DNS looks like a home line (dsl, pool, dynamic, cable, customer)
mobile_carrier
- Points: -15
- Confidence: 0.80
- Subtracts: 12
- Fires when: Mobile network (carrier data, a mobile hostname, or our CGNAT flag) with no anonymizer evidence
rdns_mobile
- Points: -10
- Confidence: 0.70
- Subtracts: 7
- Fires when: Reverse DNS looks like a mobile network
rdns_cgnat
- Points: -15
- Confidence: 0.70
- Subtracts: 10.5
- Fires when: Never fires today: our hostname classifier has no CGNAT pattern

The verified crawler discount is visible in the Googlebot lookup. datacenter_ip added 10, verified_crawler took away 28.5, and the API returned 0. The leftover minus 18.5 is discarded. It does not bank credit against a later reason, which is the right behavior: a verified range should not be able to hide a botnet listing. The 19 range files behind verified_crawler are listed, with their dates and an allowlist builder, in our verified crawler IP list; how to verify Googlebot by IP covers Google's files in detail.
We can't show you the residential discount on a live address, because publishing a lookup of a home connection would identify a household. So here is the code path in prose. The residential reasons hang off reverse DNS: our code calls a line residential only when its PTR hostname matches a home pattern such as dsl, pool or dynamic. When it does, unless the line is also flagged mobile or CGNAT, residential_isp (-17) and rdns_residential (-16) fire together, for minus 33. A mobile or CGNAT line gets mobile_carrier (-12) in place of residential_isp, so that pair is minus 28. A home line whose ISP publishes no reverse DNS, or names its hosts in a pattern we don't recognize, gets neither discount. With nothing else against it, it simply scores 0. That is the same score, so the gap only matters when other reasons fire.
The discounts also flip. If a residential line carries proxy, Tor or VPN-protocol evidence (an IPsec tag alone doesn't count), residential_isp is replaced by residential_with_threat (+21.25). A mobile line does the same with mobile_with_threat (+16). When proxy evidence is the only reason, both scale down with the age of the sighting, explained after the first group below. A discount for looking like a home is withdrawn the moment the home starts relaying traffic.
Three caps sit beside the discounts. A consumer VPN is capped at 70 (step_up) unless the address also carries abuse, spam, botnet, Tor, proxy or honeypot evidence, so a VPN alone never reaches block; a business VPN gateway is exempt and scores far lower anyway. A relay address is capped at 60, so an iCloud Private Relay egress can never reach block; we explain how to detect Private Relay and WARP traffic and why to let it through. An address our CGNAT step flags is capped at 50 unless it carries VPN, proxy, Tor, VPN-protocol, botnet or abuse evidence of its own. Proxies get a floor rather than a cap: an address we verified as a proxy, or that a third-party list reports as relaying, never scores below 26, so home-line discounts and an old sighting can't talk it down to allow.
Every detection_methods value, field by field
GeoIPHub's scoring code can name 43 reasons, and 40 of them can appear in a response today. The groups below list all 40 with their points, confidence and contribution, grouped by the question each one answers.
Is the address an anonymizer? These come from the Tor Project's exit list, providers' published server lists, our own protocol probes and a third-party proxy list. Probes cover SOCKS5, SOCKS4, HTTP CONNECT, HTTP forward, OpenVPN, IKE (v1 and v2), PPTP and L2TP. There is no WireGuard probe.
tor_exit
- Points: 45
- Confidence: 0.99
- Adds: 44.55
- What it means: On the Tor Project's official exit list, re-read every 6 hours
known_vpn_server
- Points: 40
- Confidence: 0.90
- Adds: 36
- What it means: is_vpn is true and the VPN is named: on a provider's server list, a server hostname we resolved, a public server list, a network the provider runs, or the WHOIS registration of its range
vpn_service_unattributed
- Points: 30
- Confidence: 0.85
- Adds: 25.5
- What it means: is_vpn is true because a VPN answered our protocol handshake, but no list names the provider
corporate_vpn_gateway
- Points: 15
- Confidence: 0.80
- Adds: 12
- What it means: A business VPN gateway (
vpn_typeisbusiness). It replaces the two reasons above,datacenter_with_vpn_proxydoesn't fire, and the VPN cap doesn't apply
multi_source_vpn
- Points: 30
- Confidence: 0.85
- Adds: 25.5
- What it means: A VPN whose WHOIS or RDAP record also has VPN keywords, so two sources agree
openvpn_open
- Points: 35
- Confidence: 0.90
- Adds: 31.5
- What it means: Our probe got an OpenVPN reply, or the provider's server list tags the server for OpenVPN
ipsec_open
- Points: 30
- Confidence: 0.80
- Adds: 24
- What it means: Our probe got an IKE reply, or the provider's server list tags the server for IKEv2 (the NordVPN, Surfshark and Windscribe lists do)
pptp_open
- Points: 30
- Confidence: 0.92
- Adds: 27.6
- What it means: Our probe got a PPTP reply
l2tp_open
- Points: 30
- Confidence: 0.85
- Adds: 25.5
- What it means: Our probe got an L2TP reply
wireguard_open
- Points: 35
- Confidence: 0.70
- Adds: 24.5
- What it means: The VPN provider's published server list tags this server as WireGuard. Not a handshake
socks5_open
- Points: 35
- Confidence: 0.95
- Adds: 33.25
- What it means: A SOCKS5 or SOCKS4 proxy answered our probe
http_proxy_open
- Points: 35
- Confidence: 0.95
- Adds: 33.25
- What it means: An HTTP CONNECT or HTTP forward proxy answered our probe
blocklist_proxy
- Points: 25
- Confidence: 0.70
- Adds: 17.5
- What it means: A proxy flag with no protocol detail, used only when no SOCKS, HTTP, exit, gateway or list finding explains the flag
proxy_exit_verified
- Points: 40
- Confidence: 0.90
- Adds: 36
- What it means: An address that a proxy we verified forwarded our test request out of
proxy_gateway
- Points: 35
- Confidence: 0.90
- Adds: 31.5
- What it means: An address that accepts proxy connections and hands them to other exits
proxy_observed
- Points: 35
- Confidence: 0.80
- Adds: 28
- What it means: On a third-party list of addresses seen proxying traffic, not verified by us. Kept 14 days after it leaves the list
residential_proxy
- Points: 40
- Confidence: 0.90
- Adds: 36
- What it means: A proxy we flagged ourselves, not by list alone, on a line our data calls residential or mobile, from reverse DNS, carrier data or the network the address belongs to. A coarse inference, not residential proxy detection
relay_ip
- Points: 10
- Confidence: 0.85
- Adds: 8.5
- What it means: A privacy relay egress such as iCloud Private Relay; also caps the score at 60
Proxy sightings age. The Adds value above shows the full weight. Seven reasons (socks5_open, http_proxy_open, blocklist_proxy, proxy_exit_verified, proxy_gateway, proxy_observed and residential_proxy) scale with how long ago we last saw the proxy, and so do datacenter_with_vpn_proxy, residential_with_threat and mobile_with_threat when a proxy is their only anonymizer evidence. The factor is 1 for seven days after the last sighting, then falls in a straight line to 0.5 at 30 days. Our scanner drops unconfirmed proxy rows after 14 days, so the factor usually stays at 0.85 or above.
VPN protocols count a quarter on a VPN. When is_vpn is true, openvpn_open, ipsec_open, pptp_open, l2tp_open and wireguard_open add a quarter of the value above, because the VPN reason already counts the server. On 11 November 2026 the Mullvad server's wireguard_open added 6.125, not 24.5.
Has the address misbehaved? These come from three DNSBL zones and our abuse and botnet feeds.
known_spammer
- Points: 40
- Confidence: 0.90
- Adds: 36
- What it means: Listed on zen.spamhaus.org, b.barracudacentral.org or bl.spamcop.net. Counts half on a relay address
abusive_listed
- Points: 30
- Confidence: 0.80
- Adds: 24
- What it means: Listed on our abuse feeds, or on at least two of the three DNSBL zones. Counts half on a relay address
botnet_c2
- Points: 50
- Confidence: 0.95
- Adds: 47.5
- What it means: A botnet feed lists it as a command-and-control server
botnet_infected
- Points: 35
- Confidence: 0.80
- Adds: 28
- What it means: A botnet feed lists it in any other role, such as an infected machine
Is it server infrastructure? Hosting means one of 853 curated hosting ASNs, a range from the 10 cloud providers we load, or an anycast range. Only one of the three datacenter_* reasons can fire, strongest first.
datacenter_with_vpn_proxy
- Points: 40
- Confidence: 0.90
- Adds: 36
- What it means: Hosting address with any VPN or proxy protocol evidence above, including a list's WireGuard tag. Not for a business VPN gateway
datacenter_blocklisted
- Points: 30
- Confidence: 0.75
- Adds: 22.5
- What it means: Hosting address that is also on an abuse, spam or botnet list. Counts half on a relay address
datacenter_ip
- Points: 20
- Confidence: 0.50
- Adds: 10
- What it means: Hosting address with nothing else against it
no_ptr_datacenter
- Points: 10
- Confidence: 0.50
- Adds: 5
- What it means: Hosting address with no reverse DNS in our stored records, which have gaps
whois_vpn_keywords
- Points: 20
- Confidence: 0.70
- Adds: 14
- What it means: WHOIS or RDAP record has VPN keywords, but the address is not flagged as a VPN
whois_hosting
- Points: 15
- Confidence: 0.55
- Adds: 8.25
- What it means: WHOIS or RDAP record has hosting keywords, but the address is not on our hosting lists
What does its reverse DNS say? These read the stored PTR hostname with simple patterns, which is why each one is light.
rdns_tor
- Points: 30
- Confidence: 0.80
- Adds: 24
- What it means: Hostname looks like a Tor exit
rdns_vpn
- Points: 20
- Confidence: 0.75
- Adds: 15
- What it means: Hostname contains "vpn" or a VPN brand name
rdns_proxy
- Points: 20
- Confidence: 0.70
- Adds: 14
- What it means: Hostname contains "proxy", "squid" or "socks"
rdns_hosting
- Points: 10
- Confidence: 0.60
- Adds: 6
- What it means: Hostname looks like a server or cloud host
fcrdns_mismatch
- Points: 5
- Confidence: 0.40
- Adds: 2
- What it means: The hostname does not resolve back to the address
rdns_residential
- Points: -20
- Confidence: 0.80
- Adds: -16
- What it means: Hostname looks like a home line
rdns_mobile
- Points: -10
- Confidence: 0.70
- Adds: -7
- What it means: Hostname looks like a mobile network
What kind of line is it, and is it a known crawler? The network-type reasons are always one of a pair: the discount without anonymizer evidence, or the penalty with it.
residential_isp
- Points: -20
- Confidence: 0.85
- Adds: -17
- What it means: Residential line, inferred from reverse DNS, with no anonymizer evidence
residential_with_threat
- Points: 25
- Confidence: 0.85
- Adds: 21.25
- What it means: Residential line with proxy, Tor or VPN-protocol evidence (an IPsec tag alone doesn't count)
mobile_carrier
- Points: -15
- Confidence: 0.80
- Adds: -12
- What it means: Mobile line with no anonymizer evidence
mobile_with_threat
- Points: 20
- Confidence: 0.80
- Adds: 16
- What it means: Mobile line with anonymizer evidence
verified_crawler
- Points: -30
- Confidence: 0.95
- Adds: -28.5
- What it means: Inside an official crawler IP-range file, re-fetched daily
Three reasons the code knows but never returns
Three names exist in GeoIPHub's scoring code but cannot appear in a response today, so don't build rules that wait for them:
spoofed_crawler(+42.75). Nothing in our pipeline sets it, and a lookup only sees an IP, never a User-Agent. Comparing a claimed bot name withthreat.crawlerhappens in your code.honeypot_hit(+33.25). The honeypot count is a placeholder that is always 0.rdns_cgnat(-10.5). Our hostname classifier has no CGNAT pattern.
Two names we listed here before, known_vpn_asn and shadowsocks_suspected, were deleted from the code in the 10 November 2026 release. The config file also holds 11 more weights that the lookup path never reads, so those names never reach a response either. We would rather list a dormant reason than let you discover it by waiting.
Write rules on reasons, not on the raw number
The practical reason to learn these names is that rules written on reasons survive our data gaps, and rules written on the total do not. Our block thresholds for login, signup and checkout set the bands; inside them, a few reason-level rules do most of the work:
- Treat tor_exit, socks5_open, http_proxy_open and botnet_c2 as strong on their own: each adds 33 or more on a recent sighting and comes from a list or our own probe.
- Don't act on no_ptr_datacenter, fcrdns_mismatch or any rdns_* reason alone. They carry 2 to 24 points and read stored hostnames that can be missing or stale.
- Read verified_crawler together with threat.crawler.verified and an operator that matches the bot the User-Agent claims, never threat.crawler.name. A verified crawler on a spam list, like AhrefsBot on 29 October 2026, belongs on your crawl allowlist, not in your checkout.
- Treat wireguard_open as list evidence, not a probe result. It means the provider's published list says the server runs WireGuard.
- Log the full detection_methods array with every decision. It is the audit trail a support agent or chargeback reviewer will ask for.
Here is how that looks in code, using only real fields:
const { recommended_action, detection_methods: why } = data.scoring;
const strong = ["tor_exit", "socks5_open", "http_proxy_open", "botnet_c2"];
if (why.some((m) => strong.includes(m))) return block(why);
// isCrawlPath is your own route check, not an API field.
// claimMatchesOperator compares the bot your User-Agent names with
// threat.crawler.operator; /blog/verified-crawler-ip-list has a working version.
if (why.includes("verified_crawler") && isCrawlPath(req) &&
claimMatchesOperator(req, data.threat.crawler)) return allow(why);
if (recommended_action === "block") return block(why);
if (recommended_action === "step_up") return challenge(why);
return allow(why);
Reasons also tell you where evidence comes from, which matters when a customer disputes it. known_vpn_server with vpn_confidence 95 means the provider's own list; we explain how we name the VPN provider behind an IP and what each confidence value says. For the probe side, how VPN and proxy detection works covers where the heaviest anonymizer reasons come from, and the full list of IP data fields documents the rest of the response.
Add up one score yourself
Paste any address into GeoIPHub's free IP fraud score checker, copy its reasons, and rebuild the number from the entries above. If your sum and our score disagree, one of us has a bug, and we would like to hear which.
Go deeper
Frequently Asked Questions
What does scoring.detection_methods contain?
The name of every reason that changed the score, in the order GeoIPHub's scoring code checks them, not by size. Each name has fixed points and a confidence in our scoring config, and the score is the sum of points times confidence, rounded and held between 0 and 100. GeoIPHub's code can name 43 reasons; 40 can appear in a response today.
Why does a verified crawler still get a review score?
Because in GeoIPHub's scoring verified_crawler is a weight, not an allowlist. It subtracts 28.5 points (minus 30 at 0.95 confidence). AhrefsBot at 5.39.1.230 also carried a Spamhaus listing on 2026-10-29, so the spam and blocklist reasons outweighed the discount: 35, review, down from the 64 (step up) it would score without the crawler match.
Can a score go below zero?
The weighted sum can, the score cannot. Googlebot at 66.249.66.1 summed to minus 18.5 on 2026-10-29 (datacenter_ip plus verified_crawler) and GeoIPHub's API returned 0. The negative balance does not carry over as credit for anything else.
Does wireguard_open mean GeoIPHub completed a WireGuard handshake?
No. GeoIPHub has no WireGuard probe. wireguard_open means the VPN provider's own published server list tags that server as WireGuard. Our probes cover SOCKS5, SOCKS4, HTTP CONNECT, HTTP forward, OpenVPN, IKE (v1 and v2), PPTP and L2TP, so pptp_open, l2tp_open, socks5_open and http_proxy_open are probe results, while openvpn_open and ipsec_open come from our probe or from the provider's list.
Why do I see no_ptr_datacenter on an address that has reverse DNS?
The reason reads GeoIPHub's stored reverse DNS records, not a live query, and those records have gaps. On 2026-10-29 it fired for AhrefsBot 5.39.1.230 and the Tor exit 185.220.101.1, while dig -x resolved both. It is worth 5 points, so write rules on the reasons you trust rather than on the raw count.
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.


