In large localized sites you do it by geography. (E.g. Craigslist, Trulia, Auto Trader, etc) You analyze your traffic, find two geographies that are statistical

In large localized sites you do it by geography. (E.g. Craigslist, Trulia, Auto Trader, etc) You analyze your traffic, find two geographies that are statistical neighbors and that have sufficient traffic. One is your control, the other is your experiment. Run a test with 20 control and experiment locations.