April 12, 2026 Β· 12 min read
IP Geolocation Accuracy: We Measured 2,000 Known Locations
We compared IP geolocation against 2,000 RIPE Atlas probes with known coordinates. Median error 6.4 km, mean 152.9 km β accuracy is bimodal, not merely imprecise. Full data and method included.
Ask how accurate IP geolocation is and you will get adjectives. "City-level." "Approximate." "Varies by region." We wanted numbers, so we measured it: 2,000 IP addresses whose true coordinates are known, checked against a commercial geolocation database.
The short answer is that IP geolocation is usually very good and occasionally absurd, and every average published about it hides that fact.
The finding
| Percentile | Error |
|---|---|
| 50th (median) | 6.4 km |
| 75th | 48.9 km |
| 90th | 278.7 km |
| 95th | 527.6 km |
| 99th | 2,073 km |
| Worst | 13,589 km |
The mean error was 152.9 km β twenty-four times the median. That is not a rounding artefact. It is the whole story.
| Error range | Share of lookups |
|---|---|
| Under 10 km | 58.3% |
| 10β50 km | 17.0% |
| 50β250 km | 14.0% |
| 250β1,000 km | 8.2% |
| Over 1,000 km | 2.6% |
The worst 10% of lookups carried 84% of all error. IP geolocation is not uniformly imprecise, the way a blurry photograph is. It is bimodal: most of the time it is nearly right, and a meaningful minority of the time it is somewhere else entirely.
This matters because it changes what the number means for you. A median of 6.4 km sounds like a tool you can rely on. A 10.8% chance of being wrong by more than 250 km means no single result can be trusted on its own, regardless of how good the typical case looks.
Country is nearly always right β until it isn't
Country was correct in 98.4% of lookups. That is the figure behind geo-blocking, content licensing and payment-fraud rules, and it holds up well.
The remaining 1.7% failed hard: median error of 1,066 km for those 33 probes. A wrong country is not a near miss into a neighbouring one; it is a different continent's worth of wrong. Any system that treats country as a binary trust signal is fine 98% of the time and catastrophically unfair to the rest.
Datacenter IPs are more accurate than home connections
| Connection type | Probes | Median error |
|---|---|---|
| Datacenter / hosting | 280 | 1.5 km |
| Residential | 1,720 | 7.3 km |
This inverts the common assumption, and it is the most practically useful result here.
Hosting providers publish precise facility locations, and those registrations are accurate because everyone involved benefits from them being accurate. Residential IPs are assigned from pools that move, get reassigned, and are documented against a billing office rather than a home.
For VPN users the implication is direct: your exit node geolocates better than your real connection does. The location a website sees is a precisely known building. It is simply not your building. That is the point of the VPN β but it also means "my IP location looks wrong, so I must be anonymous" is bad reasoning. It looks exactly right; it just describes somewhere else.
Where it performs worst
Countries with at least 10 probes in the sample, ranked by median error:
| Country | Probes | Median error |
|---|---|---|
| Italy | 53 | 29.1 km |
| United Kingdom | 94 | 27.2 km |
| France | 141 | 27.1 km |
| South Africa | 25 | 19.9 km |
| Netherlands | 95 | 16.4 km |
| New Zealand | 13 | 15.6 km |
| Belgium | 25 | 12.9 km |
| Poland | 30 | 10.9 km |
| Czechia | 37 | 10.6 km |
| Austria | 35 | 10.2 km |
Note what this list is not. These are dense, wealthy, extremely well-connected markets β precisely the places assumed to be well mapped. Accuracy does not track a country's internet development; it tracks how a given ISP allocates and registers address space. A national carrier that routes a whole region through one metro hub will geolocate every subscriber to that hub, however rich the country.
Method
Ground truth comes from RIPE Atlas, a global network of measurement probes. Each public probe publishes operator-declared coordinates alongside its public IPv4 address, giving thousands of addresses whose real location is known.
Critically, this means no visitor data was collected to produce these numbers β no consent flow, no storage, no tracking of anyone using this site.
- Fetch every connected, public Atlas probe with both an IPv4 address and coordinates β 13,081 at time of writing.
- Sample 2,000 with a fixed random seed, so the run is reproducible.
- Look each IP up in a commercial geolocation database (ip-api.com).
- Compute the great-circle distance between claimed and actual position.
Measurements were taken on 27 July 2026.
The script and the complete raw dataset are published: github.com/CodeWithEhtisham/WhatIsMytools β collect.py plus results.csv, one row per probe with claimed position, true position and error. Rerun it and check us.
Limitations
Stated plainly, because a study without them is marketing:
- Probe coordinates are operator-declared, not surveyed. Some are precise to the metre, others rounded to a nearby landmark. This adds noise, and probably inflates our small-error figures slightly.
- The probe population is not the internet. Atlas volunteers skew European, technical, and well-connected. Expect these results to be a best case. Typical consumer connections in under-probed regions are likely worse.
- One database, one day. These figures describe ip-api.com on 27 July 2026. Other providers differ, and all of them update continuously. The method is built to compare multiple sources β a local MaxMind GeoLite2 database can be added with one flag.
- Datacenter probes are a minority here (14%) but dominate VPN and proxy traffic, so the residential figures are the ones that describe ordinary browsing.
What actually feeds these databases
The error pattern above follows directly from how the data is assembled. Providers blend:
- RIR/WHOIS allocations β often the mailing address of an ISP, not a user
- BGP announcements and peering hints
- Latency probes and crowdsourced corrections
- Hosting vs residential classification from AS labels
A datacenter IP geolocates to its facility city because that is a documented fact. A residential IP geolocates to wherever the ISP's registration points, which may be a regional hub hundreds of kilometres from the subscriber. The tail in our data is mostly this: not database error, but an honest answer to a different question than the one being asked.
Mobile, satellite and split tunnels
Cellular users share carrier NAT pools that shift between towers. Starlink and similar constellations may register ground stations hundreds of miles away. Corporate split tunnelling shows office egress for work traffic and home IP for personal tabs β the same laptop, two different answers.
These are the conditions that produce our 2.6% of errors above 1,000 km.
If you build on IP location
- Never treat a single lookup as fact. Given a 10.8% chance of a >250 km miss, a decision that matters needs a second signal.
- Widen the radius or say "approximate" if you show a map pin. A pin implies a precision the data does not have.
- Offer manual override anywhere the answer has consequences β shipping, tax, age gating.
- Log the ASN and hosting flag, not just latitude and longitude. Knowing a result came from a datacenter tells you more than the coordinates do.
- Weight country higher than city. 98.4% versus a coin flip past 50 km is a real difference in reliability.
Cross-check your own result with our What Is My Location tool against your phone's GPS outdoors. You will land somewhere in the distribution above β most likely within a few kilometres, and occasionally in another country entirely.