Development Seed Blog

As I fired off requests to Google's geocoder API to plot several thousand points on a map for a recent project, a quick calculation showed that only about a third of the data returned valid latitude and longitude information. The data set I was dealing with was international - mapping members from nearly every country - so I knew I would have some issues, but only a third successful? I thought, come on, Google! After a few hours I developed a short list of helpful tips to follow to get a much higher success rate when geocoding data using Google or Yahoo.

I had a complete address for nearly every member I needed to geocode, but short of picking through country data sets, there's only so much latitudinal and longitudinal data out there for exact addresses. Google and Yahoo have the United States pretty well covered and an open source government data set makes it easy. However, try to geocode a house on a road that only locals know the name of and you're pretty much out of luck, unless you get that guy to send in the coordinates off his Bushnell Onix hand held GPS. Fortunately for this case, I only needed city and country coordinates.  

When I saw the one-third success rate, I knew something wasn't right. To make the actual requests on Google, I used a function in the location module (called location_geocode_us_google) that relies on passing a location as an array, which it converts into a single line address to send to Google. When Google didn't know the address, I guessed that it didn't like the entire address and was just unable to save the city/country pair after failing with the full address. Go to this Google example and toss it a full international address for somewhere in the developing world - chances are it will not be able to geocode it. Now try just the city and country - it should work.  

This was an easy guess, but before I got much further I decided to modify the Google geocoder function and record what the actual response code was for every request. There are six possible response codes.  The default behavior of the function was to check if it was 200 - which mean it was a successful request - and if it was not, don't do anything. The five other possible response codes are listed here in Google's documentation.

I tried using just city and country in the location array, and it significantly raised the success rate. However, I noticed a lot of code 603 responses, which as you can see on the above link, means the "geocode for the given address cannot be returned due to legal or contractual reasons." It turns out that Google cannot return latitude and longitude data for China or the United Kingdom, and possibly a few other countries. I didn't investigate the reasons for this, but switched over to Yahoo, which worked fine.

This experience may help make a smarter geocoding module that's trained to handle bad input data from the user and use other geocoders as fallbacks whenever a successful response is not returned. In a few weeks I should have another tale from the geocoding trail about how Yahoo's geocoder can help with split congressional districts, where the five digit zip code cannot distinguish your correct representative. Any guesses?