Copied to clipboard
← All posts
Field test · Geolocation

How Accurate Is IP Geolocation? We Checked 11,449 Addresses Against Where They Really Are

Half of them landed within 15 km of the truth. One landed 15,642 km away. Both answers came back with the same confidence, and that is the part worth understanding.

· 10 min read· 11,449 addresses measured· 181 countries

A message that turns up in our inbox again and again is some version of “your site says I am in Frankfurt and I have never been to Frankfurt.” The person writing is nearly always right. We wanted to know how often that happens and by how many kilometers, so we stopped trading anecdotes and went looking for a few thousand addresses whose real location somebody had written down.

Finding addresses that come with an answer key

The hard part of testing geolocation is not the lookup. It is knowing where the address actually is. For that we used RIPE Atlas, a measurement network made of small probes that volunteers plug into their home routers, office networks and server racks. Each host tells the project where the probe sits, and that location is public, along with the address the probe connects from.

On November 17, 2026 there were 15,528 probes online. We threw a lot of them out. 2,812 carried a tag saying their location had been filled in automatically from an IP database, which would have meant grading a database against its own homework. Another 70 had a location the network itself flags as disputed, and 1,008 had no public address to look up. That left 11,637 probes with a hand-entered location, of which 11,449 had an IPv4 address we could place. They sit in 181 countries on 3,850 different networks.

For each one we asked the same location database that our IP lookup reads, took the coordinates it returned, and measured the straight-line distance over the globe to where the host says the probe is. Nothing else. No reverse DNS, no latency triangulation, no second opinion.

97.1%placed in the right country
15kmmedian distance from the truth
57.2%within 25 km, roughly one metro area
417kmor worse for the unluckiest tenth
Fig. 1 · a random 1,600 of the 11,449
Every dot starts where we put it and travels to where it really is

Most never visibly move. The ones that do tend to cross a country, or an ocean.

  • within 25 km (912)
  • 25 to 250 km (424)
  • over 250 km (264)
Lines connect each guess to the real position. Hover or tap a dot for its country and error. The sample is random, so the color mix is the mix in the full set. Positions are rounded to the map’s resolution.
Read this before quoting any number here

The answer key is self-reported. Hosts type in a location, and some of them get it wrong or move house without updating it. The coordinates are also blurred by about a kilometer on purpose, so anything under 2 km is as close as this method can see.

The locations are public, which means any location database is free to learn from them, and some surely have. Treat our results as flattering. People who run measurement hardware for fun are also not a cross-section of the internet: 7,004 of the 11,449 are in Europe, and they are more likely than average to have a fixed line from a tidy network.

This is one database, the one behind our own API, on one day. It is not a comparison between providers and we are not claiming anyone else would do better or worse.

The median is 15 km. The average is 229.

Those two numbers describe the same test, and the gap between them is the whole story. Geolocation errors do not form a tidy bell around the right answer. 1,536 addresses (13.4 %) came back within a kilometer, which given the blurring means dead on. Nearly a third were within 5 km. Then the curve flattens and just keeps going: 423 addresses were more than 1,000 km out and 117 were more than 5,000. The record holder is a server in the United States that the database files under Singapore, 15,642 km from where it hums.

Drag the radius below. The dots are real errors from the test, laid out by how far off they were. The spacing is logarithmic, because on a linear scale everything interesting would be a smudge in the middle.

Fig. 2 · pick a radius
How close is close enough for what you are building?
57.2% inside

6,549 of 11,449 addresses were placed within 25 km of where they really are.

Left: a sample of 700 errors; the direction of each dot is arbitrary, only its distance from the center means anything. Right: the cumulative curve for all 11,449. The percentage always comes from the full set.
Show the numbers
WithinIPv4 (11,449)IPv6 (6,297)
5 km31.8 %33.3 %
10 km43.7 %46.2 %
25 km57.2 %59 %
50 km65.1 %66 %
100 km72.2 %72.3 %
250 km83.4 %82.3 %
500 km92.1 %91.1 %
1,000 km96.3 %95.1 %

All 11,449 answers were labelled city-level. The one that missed by 15,642 km included.

That is the uncomfortable finding. The database never said “I only know the country” for a single address in this set. It returned a city and a pair of coordinates every time, and nothing in the record separates the 13.4 % that were perfect from the ones that were on the wrong continent. An error you can see coming is manageable. This kind you only discover when a customer writes in about Frankfurt.

The kind of line matters more than the place

Many hosts tag their probe with what it is plugged into, which let us split the results by connection type. The differences are large enough to see on a map. Below, each circle is the median error for one kind of connection, drawn to scale. Click anywhere to move the circles to a place you know.

Fig. 3 · median error, to scale
If the address is really at the center, half the guesses land outside its circle
London Paris Frankfurt Amsterdam Brussels Manchester Lyon Zurich Birmingham Cologne Nantes Strasbourg
7 kmServer in a data center1,993 probes · 67.4 % within 25 km
22 kmHome broadband4,354 probes · 51.5 % within 25 km
74 kmMobile network151 probes · 30.5 % within 25 km
219 kmSatellite63 probes · 12.7 % within 25 km
The outline is deliberately coarse. Connection types come from tags the hosts chose; 4,774 probes had no usable tag and are left out of this figure (their median was 13 km).

Servers are easy. A machine in a rack does not move, its operator often publishes where its address blocks are deployed, and the median error for the 1,993 data-center probes was 7 km. That is also the group with the wildest misses, because a hosting company can move a block of addresses from Toronto to Singapore in an afternoon and the paperwork follows months later.

Home broadband is a coin flip at city level. Median 22 km across 4,354 probes, and only 51.5 % inside 25 km. Country was right 98.6 % of the time, the best of any group.

Phones and dishes are a different sport. The mobile median was 74 km and the satellite median 219 km. A phone’s traffic leaves the carrier’s network through one of a handful of gateways, and a satellite terminal’s leaves through a ground station that may be in the next country over. In both cases the address tells you where the gateway is. These two samples are small, 151 and 63 probes, so take the exact figures loosely. The direction is not in doubt. Probes whose hosts said they sit behind carrier-grade NAT, where hundreds of households share one public address, had a median of 59 km.

Then there is France

We expected accuracy to track wealth, or density, or how long a country has been online. It tracks none of them. It tracks how the big local networks hand out addresses.

Fig. 4 · countries with 80+ probes
Median error by country
CountryMedian errorkm≤ 25 km
New Zealand26923 %
France14435 %
Italy7138 %
United Kingdom6834 %
Denmark3847 %
Belgium2845 %
Taiwan2849 %
Netherlands2550 %
Australia2550 %
Switzerland2153 %
Iran2053 %
Norway1955 %
India1955 %
South Africa1857 %
Japan1757 %
Germany1556 %
United States1462 %
Sweden1361 %
Portugal1260 %
Austria1063 %
Greece1073 %
Finland9.565 %
Czechia9.365 %
Spain9.265 %
Ukraine7.866 %
Singapore7.795 %
Brazil7.679 %
Poland7.166 %
Russia6.776 %
Canada6.670 %
Indonesia6.164 %
Kazakhstan4.960 %
Scale runs from 0 to 280 km. Sample sizes range from 80 to 1,556 probes; hover a row for its count.
Show the numbers
CountryProbesMedian km75th pct km≤ 25 kmRight country
New Zealand106269.051822.6 %100 %
France852143.639634.9 %98.4 %
Italy29071.224538.3 %98.3 %
United Kingdom51167.721034.1 %98.2 %
Denmark8838.415246.6 %100 %
Belgium15128.35545 %96.7 %
Taiwan8028.013548.8 %96.2 %
Netherlands61125.48549.8 %94.9 %
Australia22125.215649.8 %100 %
Switzerland29021.18053.4 %98.3 %
Iran8019.589852.5 %97.5 %
Norway8718.710055.2 %100 %
India12118.625955.4 %95 %
South Africa11518.147557.4 %95.7 %
Japan21716.815856.7 %97.2 %
Germany1,50715.216055.7 %96.5 %
United States1,55613.95861.8 %98.3 %
Sweden19412.912261.3 %99 %
Portugal8711.79059.8 %98.9 %
Austria19210.110263 %97.4 %
Greece8010.13472.5 %100 %
Finland1819.54964.6 %96.7 %
Czechia2219.38264.7 %99.1 %
Spain2189.213365.1 %98.2 %
Ukraine1217.834466.1 %92.6 %
Singapore1177.71294.9 %94.9 %
Brazil1027.61979.4 %99 %
Poland1947.16365.5 %98.5 %
Russia4196.72476.4 %95.5 %
Canada2626.63369.8 %98.9 %
Indonesia946.15363.8 %96.8 %
Kazakhstan974.982459.8 %100 %

France has 852 probes in the set, the third most of any country, and a median error of 144 km. For French home connections alone it is 230 km. Germany, next door: 15 km.

Most of that gap is one company. The largest French network in the set hosts 265 probes, and the database put just 14 % of them within 25 km. Its median error was 324 km, which is most of the way across the country. The largest German network hosts 296 probes, and 80 % of those were within 25 km, at a median of 7.8 km. Same technology, same decade, same database. One of them assigns addresses out of regional pools that stay put for years. The other, as far as the outside world can tell, deals them from a national deck.

New Zealand does worst of all at 269 km, with fewer than a quarter of its 106 probes inside 25 km. The United Kingdom (68 km) and Italy (71 km) are not far behind France. At the other end sit Kazakhstan (4.9 km), Canada (6.6 km), Russia (6.7 km) and Brazil (7.6 km). Sort the chart by “best first” and the top of the list is not the one anybody would have guessed.

Where the lost ones end up

2,873 addresses were placed in the right country but more than 100 km from home. They do not scatter. They pile up.

Fig. 5 · the usual suspects
Cities the database chose for addresses that were over 100 km away
Frankfurt am Main162
Paris138
London76
Lyon35
Amsterdam34
Berlin30
Moscow29
Shiraz25
Count of in-country misses over 100 km, by the city we returned. District suffixes merged.

Frankfurt, Paris and London take 13.1 % of all the long misses between them. None of that is a coincidence. Frankfurt is where a great deal of German traffic meets the rest of the internet. Paris is where French networks are headquartered. When a database has an address block and no better clue, it files the block where the owner’s paperwork or the owner’s routers are, and for a national network both of those are in the big city.

Fig. 6 · why it happens
Nine towns, one gateway, one pin on the map
gateway to the internet where the database files all nine
From outside the network, the gateway’s address is the only one anybody sees. The distances are illustrative.

The same picture explains most of what went wrong in this test, with different labels on the box:

  • A regional gateway. A broadband provider terminates customers from a whole region on a few large routers. The address belongs to the router’s pool, and the pool is filed under the router’s city.
  • Carrier-grade NAT. When a provider runs short of IPv4 addresses, it puts many households behind one. That one address now has several hundred correct locations and the database has to pick a single wrong one.
  • A mobile core. A national carrier may have three or four exits to the internet. Your phone in a mountain village shares one with a phone in the capital.
  • A ground station. Same idea, with the last hop going through orbit.
  • A head office. Registry records list the organization that holds the block and its postal address. For a company with offices in forty cities, that says where the legal department sits.
  • A sale. IPv4 addresses are bought, sold and leased now. A block that served Toronto last year can serve Singapore this year, and every database is some number of weeks behind. The 10,000 km misses in our set mostly look like this to us.

The 2.9 % in the wrong country

331 addresses were placed in a different country from the one their host reported. The median distance for those was 1,772 km, so these are mostly not border towns filed on the wrong side of a river. They are relocated servers and overseas territories.

Really inWe saidAddresses
RéunionFrance28
GermanyUnited States13
GermanyFinland12
NetherlandsGermany8
UkraineRussia8
RussiaNetherlands7
FranceGermany7
United StatesCanada6

The top row is the best illustration of the head-office problem we have ever seen. Réunion is an island in the Indian Ocean, east of Madagascar. It is also, legally, a part of France, and its networks are run by French companies. Of the 68 probes there, the database put 28 in mainland France and 21 of those in Paris. That is a miss of about 9,369 km, to a place where the weather is noticeably different.

The Ukraine-to-Russia row is short, 8 addresses, but it is the one we would lose sleep over if we used country as a compliance control. Which is the practical point of this section: 97.1 % is a very good number for choosing a default language. It is a poor one for deciding whom the law lets you serve, because the 2.9 % are not random people. They are concentrated in exactly the places where borders are contested or administrative.

IPv6 does slightly better, which we did not expect

6,109 probes have both an IPv4 and an IPv6 address, which gives a clean comparison: same machine, same room, two lookups.

Fig. 7 · 6,109 dual-stack probes
Which of the probe’s two addresses was placed closer to it?
1,986
2,485
1,638
IPv6 closerwithin 1 km of each otherIPv4 closer
12.3 kmIPv6 median · 95th percentile 985 km
15.9 kmIPv4 median · 95th percentile 765 km

IPv6 won more often than it lost and has the lower median, 12.3 km against 15.9 km. Our best explanation is boring. IPv6 space is plentiful, so networks carve it up by region and leave it alone. Nobody is buying and selling it, and nobody needs to hide a neighborhood behind one address. The layout of the network still resembles the layout of the country.

It is not a clean win. When IPv6 is wrong it is more wrong: the 95th percentile is 985 km against 765 km for IPv4, and the right-country rate is a touch lower (96 % versus 96.6 %). There is simply less history to draw on.

What we would do with this

If you are building on IP location, ours or anyone’s, the numbers above suggest some rules.

Lean on country. It is right about 97 times in 100 and wrong in predictable places. Use it for language, currency, tax display and routing people to the nearest server. Keep a manual override somewhere visible, and do not make it the only gate on anything a regulator cares about.

Treat city as a guess you are allowed to make out loud. Pre-fill the nearest store, the weather, the shipping estimate. Then show the user what you assumed and let them change it in one tap. On our own lookup page the location carries the words “city-level estimate”. After running this test we think those words are working harder than we realized.

Never treat coordinates as a position. The latitude and longitude in a geolocation response are the center of whatever city the database chose. In this test 11,449 addresses mapped to only 6,502 distinct points. Sending a courier, a process server or an angry email to those coordinates will reach a town square.

Be careful with “location mismatch” as a fraud signal. A customer whose billing address is in Lyon and whose IP says Paris is, on this evidence, probably a customer in Lyon. A rule that fires on city disagreement will mostly catch French broadband subscribers and people on their phones.

Know what a precision field is telling you. Our API returns location.precision with the value city, region or country. It reports how specific the record is. It does not report how likely the record is to be right, and in this test every record said city. We added that field so that coordinates would not imply more than the data holds. It turns out a label can overstate things just as well as a decimal place can. The field is described in the API documentation; read it as “no finer than”.


Method, briefly. Ground truth: all RIPE Atlas probes connected on November 17, 2026 with a hand-entered location (11,637 after exclusions). Estimate: one lookup per address in the location database behind our API, no other signals. Error: great-circle distance between the two. “Right country” compares ISO country codes, so a territory with its own code counts as its own country. The probe list is public and the comparison takes a few minutes to run, so anyone can repeat this against any database, including ours next month.

If you only remember one figure, make it this one: 42.8 % of the addresses we tested were more than 25 km from where the database put them. If your product tells people what city they are in, about four in ten of them will need a way to say no.