How Accurate Is IP Geolocation? Real Numbers, Not Marketing
IP geolocation in 2026 hits about 99 percent country accuracy, 75 to 88 percent at the region or state level, and 55 to 78 percent at the city level depending on database and region. Street-level accuracy is essentially never reliable. Mobile carriers, CGNAT, satellite ISPs, and recently transferred IP blocks all degrade results. The famous Kansas farm case shows what happens when people treat default coordinates as exact.
IP geolocation is excellent at telling you which country an IP is in and steadily worse the more precision you ask for. Vendors quote optimistic city-level accuracy numbers because that is what marketing pages reward. Independent testing tells a more sober story, and the gap between the two is large enough to have caused real-world harm to real people.
The accuracy ladder
Geolocation databases improve as the search area widens. A useful mental model is to picture concentric rings around the true location and ask what fraction of lookups fall inside each ring.
| Granularity | Free databases (e.g. GeoLite2) | Paid databases (e.g. GeoIP2, IP2Location DB11) | Best case (specialist data) |
|---|---|---|---|
| Country | 97 to 99.5 percent | 99.5 to 99.9 percent | 99.9+ percent |
| Region or state | 72 to 85 percent | 82 to 92 percent | 92 to 95 percent |
| City | 55 to 78 percent | 70 to 82 percent | 80 to 87 percent |
| Within 25 km | 60 to 80 percent | 78 to 88 percent | 85 to 92 percent |
| Within 5 km | 20 to 45 percent | 40 to 60 percent | 55 to 70 percent |
| Street level | Essentially never | Essentially never | Rare and unreliable |
These ranges align with the long-standing academic measurements: Poese et al., IMC 2011 still hold up directionally, and more recent comparisons by independent reviewers continue to land in the same bands. MaxMind publishes its own accuracy comparison page with numbers a few percentage points higher than third-party tests, which is normal for vendor self-reporting.
Why country accuracy is so high
IP allocations are issued by the Internet Assigned Numbers Authority to five Regional Internet Registries (ARIN, RIPE NCC, APNIC, LACNIC, AFRINIC), which then assign blocks to specific organisations within defined service regions. The country is essentially metadata baked into the allocation chain. Even a stale or sloppy database gets the country right almost all of the time, because changing country requires a formal IP transfer.
The exceptions tend to involve:
- Multinational corporate networks where a single block serves offices in several countries.
- Satellite ISPs like Starlink whose ground stations and customer geography can diverge.
- Cloud providers with anycast or migrated capacity.
- Recently transferred IP blocks where the database has not yet caught up.
Why city accuracy is so much worse
Country sits in metadata; city has to be inferred. Geolocation vendors stitch together several signals:
- WHOIS and registry data: organisational address of the IP block owner, often a head office rather than where the IPs are actually used.
- Reverse DNS hostnames: city codes embedded in PTRs like
ams-1.example.net. See reverse DNS explained for the mechanics. - BGP routing data: which networks announce the prefix, and from where.
- Active probing: latency measurements from known points to triangulate.
- User feedback and crowdsourced data: corrections submitted by users, plus anonymized signals from app SDKs and Wi-Fi positioning data (the most controversial input).
Every one of those signals can be noisy or stale. A datacentre in Frankfurt may receive a block that historically lived in Amsterdam. A mobile carrier may concentrate national traffic through one or two gateway cities. CGNAT means thousands of subscribers across a region share one public IP. The result is that even a best-in-class commercial database is wrong at the city level a meaningful fraction of the time.
⚠️ Vendors usually report city accuracy as the percentage of lookups where the returned city matches the true city. They do not always tell you the distance error when it is wrong. A wrong result can be in the next town or 400 km away.
The Kansas farm case
The most cited real-world example of bad IP geolocation hurting real people is the MaxMind Kansas farm incident, documented in detail by Kashmir Hill at Fusion in a 2016 investigation. The short version:
- For years, MaxMind's default coordinate for the United States was 38.0000, -97.0000.
- That coordinate happens to sit on a working farm in Potwin, Kansas owned by the Taylor family.
- When databases could not place an IP precisely, they returned that default.
- Over time, more than 600 million IP addresses came to resolve to that single point.
- The family received a steady stream of FBI visits, IRS investigators, suicide-prevention welfare checks, and accusations from people whose property had been stolen by someone using an IP that geolocated to the farm.
MaxMind eventually moved the default coordinate to a body of water and added explicit accuracy radii to its data. The case is a permanent reminder that geolocation outputs are best treated as probability distributions rather than facts. The lawsuit settlement and the ongoing harm took years to resolve.
Where geolocation degrades fastest
Mobile networks
Mobile carriers route subscriber traffic through a small number of national or regional gateways. An IP that geolocates to Paris may belong to a Marseille user; an IP in São Paulo may serve customers across all of Brazil. Some carriers publish more granular data, most do not.
CGNAT
Carrier-grade NAT consolidates many subscribers behind one public IP. By definition the public IP cannot resolve to a single subscriber's city. The shift to CGNAT is driven by IPv4 scarcity, covered in IPv4 vs IPv6, and it is one of the main reasons city-level accuracy has flatlined despite vendor improvements.
Satellite ISPs
Starlink and other low-earth-orbit constellations break the historical assumption that an IP block sits in one place. Customer IPs may route through ground stations in a different country than the user. Geolocation vendors are still catching up.
VPNs and proxies
By design, a VPN exit node geolocates to the country and city of the exit, not the user. That is the entire point. The mechanics are explained in how a VPN works, and the cat-and-mouse game with detection is in how streaming services detect VPN traffic.
Recent IP transfers
A /24 block that was used in Tokyo last year may now serve a hosting customer in Dublin. Databases update on different schedules, ranging from days (commercial paid) to a full month (free Lite editions). A block in transition produces wrong answers across multiple providers simultaneously.
📌 If you are building a product that uses IP geolocation, query at least two providers and treat disagreement as a signal that the answer is uncertain. Disagreement rates at the city level run 20 to 30 percent.
Vendor claims versus reality
MaxMind's GeoIP2 City product currently claims 80 percent accuracy at the city level within the US. Independent benchmarks usually find it lands in the 65 to 75 percent range against control sets that include rural and mobile IPs, which is closer to honest. IPinfo, IP2Location, and db-ip make similar marketing claims and land in similar real-world ranges. The full comparison sits in our geolocation API roundup.
The gap is not deceitful so much as definitional. Vendors test on validation sets weighted toward dense urban broadband (where they shine). Real production traffic includes mobile, CGNAT, datacentre, satellite, and corporate VPN traffic in a mix that depresses the headline number.
When the accuracy matters
Use cases vary in how much error they tolerate.
- Currency selection in ecommerce: country-level accuracy is plenty.
- Localization of language: usually country, sometimes region; accuracy is fine.
- Content geo-restrictions: country accuracy works for licensing. See what geo-blocked means and the legal status of bypassing restrictions.
- Fraud screening: country and ISP type matter more than city precision.
- Delivery of local results: city accuracy of 70 to 80 percent is usually enough to be useful and not enough to be reliable.
- Emergency services or law enforcement: not appropriate. IP geolocation has been involved in serious wrongful enforcement actions and should never be the sole basis for action against an individual.
How to test accuracy for your own population
If your service serves a known geographic area, you can build a quick sanity check:
- Collect a sample of recent user sessions with both their reported geolocation (from your IP database) and a self-declared location field (postal code, city, or selected during signup).
- Compute the rate of agreement at country, region, and city level.
- Plot the distance error distribution for the city-level disagreements.
- Repeat with a second provider to gauge variance.
You will almost certainly find your numbers landing in the ranges in the first table of this article. If they are dramatically better, your user base is concentrated in a region where the database shines. If they are dramatically worse, you probably have a lot of mobile or VPN traffic.
The honest takeaway
IP geolocation is a useful approximation, not ground truth. Country: trust it. Region: usually right. City: right most of the time, wrong often enough to design around. Street: do not. Anyone selling you precise geolocation from an IP address alone is overpromising. Pair it with explicit user input, browser geolocation (with consent), or transaction-level signals whenever the cost of being wrong is more than mild annoyance. The closely related questions of what an IP genuinely reveals and whether someone can find you with it rest on exactly the same accuracy ceiling.
Frequently asked questions
How accurate is IP geolocation at the city level?
Free databases like MaxMind GeoLite2 hit roughly 55 to 78 percent city-level accuracy depending on region; paid commercial databases reach 70 to 82 percent. Accuracy is highest in dense urban areas of North America, Western Europe, and parts of East Asia. It is lowest on mobile networks, in regions with heavy CGNAT, and on recently transferred IP blocks. No public database reliably exceeds 85 percent at the city level globally.
Can IP geolocation find my exact address?
No. Street-level accuracy is essentially never reliable from an IP address. Geolocation operates on city or postal-area approximations at best. The famous Kashmir Hill investigation of the Potwin, Kansas farm showed exactly what happens when default coordinates get treated as precise locations: more than 600 million IPs resolved to one rural property, drawing years of mistaken FBI visits and harassment. Treat any street-level claim from an IP database as wrong by default.
Why do free IP geolocation databases get it wrong so often?
Three main reasons. First, they rely on signals (WHOIS, BGP, reverse DNS) that are themselves noisy and slow to update. Second, they update less frequently than paid feeds (monthly versus daily or weekly), so recently transferred IP blocks return stale data. Third, modern realities like CGNAT, mobile gateways, and satellite ISPs concentrate many users behind few public IPs, making city-level resolution structurally impossible regardless of how good the database is.