A new method developed by Wang et al. claims to be able to geolocate IP addresses much (they claim 50 times) more accurately than has been previously feasible. The new method uses a three-tier approach: 1) sending a data packet and measuring how long it takes to bounce back, indicating an imprecise distance; 2) doing the same with all institutions with known geolocation (for example, universities or businesses) within 200 km distance of the first guess and comparison of bounce-back-times to refine the estimated location; 3) doing a similar refinement to 2) at a finer resolution.
In order to get the accuracy of the method by Wang et al. (few hundred metres or less), people presently need to volunteer their location. However, the new method is client-independent, that means no opt-in is needed to rather accurately detect the location of the computer you are using… Is that an evil advertiser I hear chuckling?
More details can be found under the source links or via the symposium where Wang et al. presented their findings (abstract and paper; video needs credentials)
(via PopSci and NewScientist)