Summary

  • A study of RIPE Atlas history found 731 probes whose reported locations failed a conservative speed-of-Internet test at least once from 2019 through April 2026.
  • After researchers and RIPE NCC contacted operators about 258 active cases, 93 of the 103 operators who replied updated their reported locations.

The map pin was in Germany. The packets were answering from Southeast Asia. For more than a year, one RIPE Atlas probe kept telling the database it lived near the Rhine while its fastest replies said otherwise. In April 2023, its operator moved the pin to Brunei. The route had been giving away the plot all along.

That is one case in a study of the RIPE Atlas measurement network, and it points to an unglamorous fact about Internet maps: the coordinates are part of the measurement. A probe’s location can help researchers infer where an address, service or network sits. Give the probe the wrong city and a clean set of latency readings can produce a polished answer about the wrong city.

Katherine Izhikevich and co-authors examined historical measurements from January 2019 to April 2026. Their paper identifies 731 unique probes—2.4% of the probes in the study—that at some point appeared to respond from a place inconsistent with the location recorded in RIPE Atlas. In April 2026, 155 were still violating the test. The monthly count rose from 22 in January 2019 to 155 seven years later.

The researchers did not try to guess where a probe “really” lived from one suspicious ping. RIPE Atlas already collects built-in DNS measurements: connected probes query each of the 13 root-server identities every four minutes. The team reused the historical record, sampling three times a day on the first day of each month. It kept the minimum round-trip time for each probe and root-server instance. A packet cannot cross a physical distance faster than light can travel through fibre; if the measured minimum beat the theoretical minimum from the probe’s reported coordinates, the pin had a problem.

The test was deliberately difficult to fail. A probe had to violate the physical bound against at least two root-server identities run by different operators. The study built in a 100-kilometre location margin and excluded Starlink probes, whose satellite links can beat a terrestrial-fibre bound. The researchers also checked root-server locations independently. The result is a floor, not a census of every stale or inaccurate coordinate.

Then came the part that turned a paper into an operational correction. From July 2025 through April 2026, the researchers reported 258 currently violating probes to RIPE NCC. RIPE Atlas notified the operators and let them accept or dispute the proposed correction. Of 103 operators who responded, 93 updated their reported location and 10 disputed the proposed correction. In 52% of location updates, the probe had already violated the test for at least 13 weeks. Geography has a remarkably long ticket queue.

Some pins were simply late. A host moved a probe and forgot to update its location. Others appear to have been entered incorrectly from the start. Hosted servers add another wrinkle: the person who installed a virtual probe may not know where the rented machine physically sits. Among the ten operators who denied the proposed correction, four said their probe ran on a third-party virtual server. Their country-level claim could fit the evidence while the city pin still did not.

The downstream cost is larger than a misplaced dot. RIPE NCC’s IPmap service uses Atlas latency to estimate the location of IP addresses. In one recent year, the researchers found 28 likely mislocated probes among 2,800 used by IPmap; those probes could affect 1,125 city-level estimates among 98,000 candidate addresses. In an earlier study, 33 suspect probes could touch 4.3% of 8,000 candidate IPs. A bad vantage point does not stay in its own box. It lends its coordinates to every inference built on top of it.

RIPE Atlas already has a dispute button and system tags for disputed and resolved locations. The study gave that machinery a steady stream of measurement-backed reports. That is the useful model: keep the network broad, make the reported metadata challengeable, and bring the evidence to the person who can correct it. A volunteer map does not need every pin to be perfect on day one. It needs a way to notice when physics has sent the pin a forwarding address.

Sources: Izhikevich et al., “Rooting Out Incorrect RIPE Atlas Probe Geolocations,” arXiv:2409.19109v4; RIPE Atlas documentation on built-in measurements, probe geolocation and probe tags.