Design Testing

November 19th, 2011
experiment, tech, work
One of the things I really like about working on websites is that we can run real experiments. If we have a change we're considering making, we can have half our users see the new version while the other half see the old version, and we can see which one performs better. [1] These are randomized, controlled, double blind trials, with no publication bias issues, and a successful result means a better version of our site that we can start using immediately. After years in school where running a proper experiment meant weeks of careful experiment design, laborious data collection, compromises in experimental procedure for the sake of practicality, insufficient sample sizes, poor generalization, and unclear usefulness, this is really satisfying.

One humbling aspect is that I've realized I'm not very good at predicting whether a change will help. None of us are. When we test new designs, sometimes they work well and other times they don't. [2] For an example of this, consider two redesigns from the early days of our daily deals website. The first is an email design, the second is a site design:

Old:
New:
Old:
New:
One of these was a 14% improvement, the other a 27% degradation. Can you tell which was which?


[1] Not all websites have an obvious metric for "performs better". For example, how does wikipedia know if a site change improves things for their users? (More edits? Better edits? More time reading? Less time?) We're generally trying to sell things, however, so we can mostly just look at the fraction of users who advance to the next step in the sales process.

[2] This really shows the value of testing: if we just made every change we thought was good, we wouldn't improve anywhere near as much as just adopting the changes that help.

Referenced in:

Comment via: google plus, facebook, substack

Recent posts on blogs I like:

Living in the AI world

I. I know lots of people who believe that superintelligent AI is going to be invented maybe next year, maybe by 2030, and that when the superintelligent AI comes maybe everyone is going to die.

via Thing of Things November 18, 2026

I did a meditation and it sort of changed my personality?!

a trip report

via benkuhn.net November 8, 2026

Test/verification technique effectiveness with Astra

When we looked at how well agents used test/verification techniques in August1, we saw that GPT-5.6 Sol generally did what you might call a caricature of various techniques when asked to use a technique or library such as TDD, QuickCheck, Lean 4, etc.; no…

via Posts on November 2, 2026

more     (via openring)