Two weeks ago I was on vacation in San Diego. Of course I took my MacBook Pro, particularly since the rental had great wifi. So I was able to check in at work, maintain my FriendFeed addiction, locate the nearest Peet’s and do research on attractions and restaurants.
Upon returning from vacation I caught up on RSS via Google Reader. In fact, I was searching for new feeds and clicked on the ‘browse for stuff’ option. Now, I can’t say Google recommendations have been that great, but it’s easy, it refreshes quickly and it has provided some decent matches.
Geolocated Google Reader Recommendations?
Much to my surprise I saw a number of recommendations with a San Diego theme. Simple GeoIP was my first thought. But that didn’t explain the fact that many of the recommendations were related to San Diego and food. (Sadly, I declined these before I put two and two together and took a screen capture.) The only other food related blog I maintain is the fantastic TasteSpotting. But that’s where it starts and ends.
Then it dawned on me. I’d been searching and surfing San Diego restaurants! Sure, I used Yelp and OpenTable, but I searched Google (and Google Images) for restaurants with the best views. In addition, I would click through to the restaurant’s website to see the menu.
With a little research I confirmed that recommendations are based on Web History (emphasis is mine.)
Confused Google Reader Recommendations
The Google Reader recommendations algorithm is easily misled by a vacation or a spate of searches on a specific topic. Together, as in my case, and it’s even worse.
This isn’t a new problem.
Marketers have long had issues with this type of behavior. Buy a baby shower gift and you might suddenly be presented with a host of baby products. Get a Gilmore Girls DVD set for your mom and you wind up getting a promotion for The Sisterhood of the Traveling Pants 2!
The complexity of trying to identify and exclude these ‘non-standard’ signals often make recommendation engines ineffective or just plain wrong.
Simple Google Reader Recommendations
As tempting as it is to use web history and location to generate recommendations it might be better to simply rely on feed history and collaborative filtering of aggregated subscriptions. Just because you can doesn’t mean you should.
At a minimum, it’s time for Google Reader to turn the dial down on web history and location so recommendations don’t suffer from topical whiplash.