My personal understanding of the location-based limitation is that ARKit is using mapping data to localize the position of the user, and that mapping data has only be prepared for a certain number of locations. By using what I understand to be the "Look Around" images, which you may be familiar with from Apple Maps, ARKit can locate a user not just by using GPS/location services, but can couple that with the iOS device's sensors (for direction/altitude) and image-based mapping for precise location and distance fro a given point. In places that are very condensed (a great example being New York City or San Francisco), using using image-based mapping data allows for exact positioning and creating AR experiences that interact with the environment.
I think a great example of this is seen in the
WWDC 2020 session on ARKit. The engineer uses the Ferry Building in San Francisco as a point of interest and builds an AR experience using the GPS coordinates of the Ferry Building, as well as the altitude. Without localizing the user's location, it might be tough to pinpoint exactly how far from the Ferry Building the user is; San Francisco is a dense city, and the user could be ten feet from the building, twenty feet from the building, or across the street from the building. Effectively matching what the ARKit camera sees, the localization helps to ensure the experience in consistent across all users.
In short; it seems the list of variable cities is growing, but is not available everywhere yet.
