So, here are the most densely Wikipedia-article-populated parts of the world, at several scales.
A diameter of 1 km finds very compact city centers, or particularly dense areas of larger cities. (Click on a city's name to recenter the map.)
Honorable mentions for the top North American and South American locales:
Moving up to a 10-km diameter gives us a small city's metropolitan area, or a large city's urban core.
Honorable mentions for the top locales on continents other than Europe and North America:
Finally, an 100-km diameter is the size of a large metropolitan area, or an urban agglomeration.
Here we get a few urban agglomerations that didn't make the previous lists individually, but do when combined, most prominently two European agglomerations that geographers have already given names to: the Rhine-Ruhr region of Germany, and the Randstad of the Netherlands.
In addition, two odd entries cover large swathes of rural Catalonia. While these each have an anchor point (Andorra and Barcelona, respectively), a large number of the points in the area are actually in rural Catalonia, and the articles only exist on the Catalan-language Wikipedia, apparently a result of an effort to import articles on nearly every village and geographical feature in the region.
The simplest way of determining the areas of greatest coverage would have been to take the latitude/longitude coordinates and bin them with some bin size. For example, with a bin size of 1 degree in each direction, the coordinate 40.6° N, 22.9° E would go into a bin bounded by 40-41° N and 22-23° E. Then we'd just sort the bin counts.
This has two problems though. The bins don't have the same area in different parts of the earth, since lines of longitude converge towards the poles. The difference is enough to be fairly significant when comparing, say, Cairo versus Copenhagen (this effect is what makes the familiar Mercator map projection distort land sizes towards the poles). Also, the boundaries between bins are arbitrary, so clusters of points that happen to straddle a boundary will get diluted between two (or even possibly four) different bins.
There are possible hacks to mitigate the problems. We could address the first by making the bin width in degrees vary with latitude. To fix the second, we could scan a window across the globe, moving it in increments smaller than the actual size of the window. The main advantage of the binning method, though, was simplicity and ease of computation. Once we're scanning a non-fixed-size window across the world, it's about as easy to do something better directly.
A more direct definition of the most-article-populated area: what is the point on the globe that has the most coordinates within a distance of n kilometers from the point? Since the earth is roughly a sphere (close enough for our purposes), by distance we mean great-circle distance.
Technical aside: We can even retain most of the computational efficiency of the latitude-longitude binning approach with an optimization: For a given candidate point, we can compute a lat-lon bounding box that lets us quickly exclude most points, and which also lets us use spatial indexes based on rectangular regions.
That leaves how to choose the candidate points. Conceptually, we want to check every point on the earth's surface, all infinity of them. An approximation is to scan a point across the globe with some step size. This spends a lot of time determining that empty circles are empty, though, especially if we use a small step size. A heuristic improvement we'll use is to use each coordinate in our data set as a candidate for a circle center; that gives us an adaptive step size that tries more finely spaced candidates in the denser parts of the data set, and more coursely jumps through the less-dense parts (trying nothing at all in the parts that have no data points). Finally, we remove circles that overlap with a circle that had a higher count; otherwise the top 10 circles would all be slight offsets of each other.
Data from Wikipedia-World as of the March 2011 database dump. Further preprocessed to only include one coordinate per article; otherwise the results are dominated by list articles that include hundreds of coordinates each. Maps from OpenStreetMap, overlays using Polymaps, and heavy lifting by Perl.