Ask three different databases which country 90.68.0.0/16 belongs to and you get two answers. The registry file says France. The routing table says Spain. A geolocation lookup says Spain too, and names a suburb of Madrid. None of them is wrong. They are answering three different questions, and most IP data consumers never notice which one they asked.
This morning at RIPE 93 in Sofia, in a lightning talk at 10:10 local time, Jim Cowie (Internet History Initiative, and a fellow at Harvard's Berkman Klein Center) and Emile Aben of the RIPE NCC put a number on the gap between the first two answers. For every IPv4 and IPv6 prefix in the routing table they compared the country code in the NRO delegated statistics with the country code of the ASN that originates it in RIPE RIS. The two agree 88.1% of the time. For the remaining 11.9%, a model built from distance, shared borders, shared language, same RIR, colonial history and sibling territories explains only about 18% of the difference. The rest is business: multinational LIRs, cloud companies registered through Singapore, and leased legacy space.
One prefix, three countries, and the files each one comes from
The talk's worked example is worth running yourself because every step is public. Start with the NRO delegated statistics, the daily file in which the five RIRs list every allocation with a two-letter code. The record covering 90.68.0.0 is ripencc|FR|ipv4|90.64.0.0|524288|.... The RIPE Database agrees at the allocation level: the parent object 90.0.0.0 - 90.71.255.255 is FR-TELECOM-20060302, country FR, allocated to France Télécom, now Orange, in March 2006. (The deck quotes the neighboring 90.0.0.0 record with 4,194,304 addresses, which actually stops at 90.63.255.255. The /13 record that does cover the example is also FR, so nothing changes.)
Now look at the routing table. RIS collector RRC21 at France-IX in Paris sees the prefix with the AS path 14907 5511 12479: Wikimedia's view through Opentransit Orange to the origin, AS12479. The delegated stats record for that ASN reads ripencc|ES|asn|12479|..., and the holder is Orange Espagne SA. So the origination country is ES.
Finally, ask where the addresses are actually used. We ran 90.68.1.1 through the Ipregistry API today and got Spain, region Madrid, city Pozuelo de Alarcón, connection type isp, route 90.68.0.0/16, organization Orange Espagne SA. That is a measured answer, built from latency, operator geofeeds and corrections, and it has no reason to know or care that the parent allocation was signed in Paris.
There is a fourth wrinkle the slide does not show. Inside the RIPE Database, the assignment object for 90.68.0.0/16 itself (netname UNI2-NET, "Orange Spain Network", status ASSIGNED PA) carries country ES, while its parent allocation carries FR. Which registry answer you get depends on whether you read the delegated stats file or the most specific WHOIS object. Our own ASN lookups report the country of the address block itself, so all 6,784 IPv4 prefixes we list for AS12479 show a registry country of ES, and someone parsing the NRO file gets FR for the same block.
The RIPE Database documentation is refreshingly blunt about why. The country: attribute of an inetnum, it says, "has never been specified what this country represents. It could be the location of the head office of a multi-national company, where the server centre is based, or the home of the End User. Therefore, it cannot be used in any reliable way to map IP addresses to countries." The statistics exchange format is equally clear that cc is "the organization to which the allocation or assignment was made" and not where the resources are used. The registries have been telling us this for nearly twenty years. The industry kept shipping "IP to country" products built on the field anyway.
Geography explains a fifth of the gap. Business explains the rest
The interesting part of the talk is the modeling. Cowie and Aben took the 11.9% residual and asked what the simplest explanation would be if the internet followed the same rules as trade. They fit a Poisson regression across 188 countries and 35,156 directed country pairs, the "gravity model" economists use for exports, with six variables: log distance between country centroids, a shared land border, membership of the same RIR, a former colonial relationship in either direction, a shared language, and sibling territories. It is the model that says a Belgian registration routed from the Netherlands is unremarkable.
It explains about 18% of the cross-border origination. That is a real effect, and a small one.
What is left over clusters into corridors that no map predicts. Prefixes registered in France and originated from Spain run at 12 times what the gravity model expects, because Orange acts as the LIR for its regional subsidiaries and the Spanish network has its own ASN. Singapore to China runs at 8.8 times expected: Alibaba and Tencent hold address space through Singapore entities. Seychelles to Hong Kong runs at 3.7 times expected, and the United States to Romania at 1.9 times, which the speakers attribute to a Romanian hoster leasing US legacy space.
To soak up those corridors the authors added a dozen or so single-country variables, among them GDP per capita, a National Internet Registry, the Financial Secrecy Index, special administrative region status, V-Dem's autocracy classification, US sanctions, the number of RIPE LIRs, the share of pre-1997 legacy IPv4, the inbound transfer rate, PeeringDB facility count and RIR region. With those in, the authors say six mechanisms explain 99.6% of cross-border origination. Laid over the whole table, the final model sorts prefixes into six cases: the 88.1% that agree, then cross-border cases of 7.13%, 1.95%, 1.61%, 0.63% and 0.15%.
Their conclusion is the one to take away: these are not errors. The legal registry country and the country of use are not required to agree. Nobody is going to make them, and if anything the trend runs the other way. 29.7 million IPv4 addresses changed hands in the first half of 2026, and a transferred block keeps its original allocation date in the stats file while its country code follows the paperwork, not the routers. Leasing rarely touches the registry at all.
Measurement says the registries are mostly right, at a coarser grain
Two weeks before RIPE 93, at IMC 2026 in Karlsruhe, Robert Beverly, Amreesh Phokeer and Oliver Gasser presented WHEREIS, which attacks the third country directly. They took over 8.2 million IPv4 and 0.8 million IPv6 WHOIS registrations across the five RIRs, sampled 62,084 IPv4 and 43,000 IPv6 prefixes with responsive addresses, and pinged each from 20 RIPE Atlas probes. Any probe whose minimum round-trip time is too short for the registered region, at two thirds of the speed of light, rules that region out.
At that granularity the registry looks good: 98.2% of IPv4 prefixes and 97.7% of IPv6 prefixes are "fully geo-consistent", meaning the allocating RIR, the holder's country and the measured location all fall in the same RIR region. The headline hides the variation. More than 99.6% of RIPE NCC prefixes pass, and 98.8% of ARIN's, but only 92.2% of AFRINIC's, and 14.7% of AFRINIC IPv4 prefixes are held by organizations outside the region. AFRINIC staff reviewed 106 of the flagged prefixes and called 77 of them known and unacceptable use.
The two studies are not in conflict, although 11.9% and 2% look like they should be. Cowie and Aben count country mismatches across everything in the routing table, including the perfectly ordinary case of a French parent allocation routed by a Spanish subsidiary. WHEREIS checks consistency only at the grain of whole RIR regions, so Paris and Madrid are the same place to it. The one tests the paperwork against the routing, the other tests the paperwork against physics, and physics is far more forgiving at continent scale.
WHEREIS also checked three commercial geolocation databases against the prefixes it found physically out of region. IPinfo caught 65% to 92% of them depending on the RIR, MaxMind 19% to 47%, DB-IP 7% to 38%. The authors suggest the laggards may still lean on WHOIS location fields. To be fair, Gasser works at IPinfo, and the paper says so. The gap is still a useful warning: a database that copies the registry country will be confidently wrong on exactly the prefixes where it matters.
Which country your IP data should use for each decision
Once you accept that there are three countries, the question for any product is which one its rule actually needs. For content licensing, tax on digital services, age-verification laws and most fraud scoring, the question is where the person is, so the answer is the measured country, read together with the VPN, proxy and relay flags that tell you whether the measurement is of the user or of an exit node. Registering Orange Spain's customers behind 90.68.0.0/16 as French because the parent allocation says FR would be a plain error.
Sanctions screening is the case where the other two countries earn their place. Programs that target an entity, not a territory, care who holds the resource and who operates the network, and a prefix registered to a sanctioned-jurisdiction holder and routed by an ASN in a third country is precisely the pattern the registry and routing countries expose and measurement hides. Use all three, and log which one fired.
For fraud rules, a disagreement between the three is a signal and never a verdict. Roughly one routed prefix in eight has a registry country that differs from its origin ASN country, and most of those are a cloud company's corporate structure or an Orange Spain, not a criminal. Score the mismatch, weight it by ASN type (a hosting ASN originating space registered half a world away deserves more suspicion than an incumbent's subsidiary), and let the measured location carry the decision.
If you only store one country column, store the measured one. It is the only one of the three that describes the user rather than the paperwork, and country-level GeoIP is right more than 99% of the time.
Ipregistry returns all three. An IP lookup gives the measured location in location.country, with the originating ASN, its organization and the announced route alongside. An ASN lookup adds the holder's legal country and, for every announced prefix, the registry country and RIR, with the documentation stating in plain words that the legal country may not correspond to where the addresses are used. You can try the API with 20,000 free lookups to get started.
Keep users in the loop Ship release notes that get read. Try Noticeable