Architecture update: leaner, meaner and faster A/B tests
visualwebsiteoptimizer.com
visualwebsiteoptimizer.com
Um, no.
A/B testing is a balancing act between theory and pragmatism. Everyone wants to get the perfect answer, but we need to get it fast enough to make useful business decisions. When your website doesn't have that much traffic this means you need to make compromises. Running multiple tests at a time on a page is such a compromise.
The one time you absolutely cannot justify multiple tests at the same time due to predictable interaction effects is when you've assigned users a sequence and use the sequence number mod n to figure out which version they get. That's bad, so don't do that. Even if the tests are on different pages you'll get interaction effects. Assign people to each test randomly instead.
Yes, running multiple tests is risky and we discourage it because even if you gain speed, the interaction effects may dissolve any results you get.
Again, your last point is excellent. Assigning user ids and then choosing variations dependent on it is a poor strategy. In VWO, we assign variations randomly and then remember (via a cookie) which variation did the user see.
Ben has IIRC previously written, and I agree, that interaction effects are typically minimal in practice unless you have catastrophic implementation choices. Since you're taking care of the implementation and it isn't catastrophic, you can safely ignore them. Don't advise your users against using your software in a manner likely to increase the value they get out of it. Interaction effects are worth an asterix in the documentation, but not much more.