Out of the cesspool and into the sewer: A/B testing trap(blog.asmartbear.com)
blog.asmartbear.com
Out of the cesspool and into the sewer: A/B testing trap
http://blog.asmartbear.com/local-minimum.html?utm_source=feedburner&utm_medium=feed&utm_campaign=Feed%3A+blogspot%2Fsmartbear+%28A+Smart+Bear%3A+Startups+%2B+Marketing+%2B+Geekery%29
6 comments
I'm not sure the big bang is always testable. The question I'd ask myself is, "How satisfied are we with the velocity of our business, given my goals?" The goals here are important. One person's goals might be around lifestyle, low-effort income, etc. Another person's goals might be around big $ liquidity or getting enough traction/growth to attract angel funding.
If growth/success is not measuring up to one's goals, you have two choices. You can A/B test your way up or you can do something dramatic (hire a salesguy, pivot your product or positioning, etc). I think a lot of startups with big dreams and crappy growth dive in with A/B tests when they should be figuring out how to change how people talk/think about their product(s).
You're right about being expensive. The scarcest fuel for most teams is optimism and confidence. One or two failed pivots can be devastating.
If growth/success is not measuring up to one's goals, you have two choices. You can A/B test your way up or you can do something dramatic (hire a salesguy, pivot your product or positioning, etc). I think a lot of startups with big dreams and crappy growth dive in with A/B tests when they should be figuring out how to change how people talk/think about their product(s).
You're right about being expensive. The scarcest fuel for most teams is optimism and confidence. One or two failed pivots can be devastating.
working backwards from your goals is the best advice i've heard in a while. in the particular case where you've raised a lot of outside capital it can be pretty easy to forget how big you are supposed to get and how quickly you are supposed to get there to put yourself solidly in the success category.
+1 on dramatic!
+1 on dramatic!
The main issue I battle with with A/B testing is backing out far enough to know whether the change is related to the edits made or simply a seasonal/daily/hourly/freak change.
I manage CPC campaigns for clients, and often times I make A/B changes in the keywords/bids/ads and wait 30-60-90 days before measuring and making changes based on the findings. And even with 90 days and tens of thousands of visits, I find that sometimes my scope is too limited and I get nailed by seasonal/yearly changes that the numbers didn't reveal.
A/B testing in 5K chunks works if everything else in the universe remains constant, but here in the real world that's never the case.
The worst is when you make a change, measure results and implement said change, and then something related to the economy (or even Weather) horks your CAC and suddenly the client isn't so thrilled with your amazing A/B improvements.
I manage CPC campaigns for clients, and often times I make A/B changes in the keywords/bids/ads and wait 30-60-90 days before measuring and making changes based on the findings. And even with 90 days and tens of thousands of visits, I find that sometimes my scope is too limited and I get nailed by seasonal/yearly changes that the numbers didn't reveal.
A/B testing in 5K chunks works if everything else in the universe remains constant, but here in the real world that's never the case.
The worst is when you make a change, measure results and implement said change, and then something related to the economy (or even Weather) horks your CAC and suddenly the client isn't so thrilled with your amazing A/B improvements.
Here's a thought: apply successive A/B testing in steps similarly to the Metropolis-Hastings algorithm. The core sys that for each new test (each new "jump") you evaluate the probability ratio a = p_new/p_old where p_x is the probability that x is successful. If a is greater than 1, go with the new, if it's less than 1, still go with the new a% of the time.
By implementing parts of this algorithm you might be able to generate a random walk scenario that is local minima resistant.
By implementing parts of this algorithm you might be able to generate a random walk scenario that is local minima resistant.
Instead of just running A versus incremental-change A2, also run a B version that's radically different from A. Thus you reap the straightforward benefits of incremental improvements while also searching for something that could radically improve your revenue.
A very powerful concept that you don't see done very often.
A very powerful concept that you don't see done very often.
It is worth looking at "Simulated Annealing":
http://en.wikipedia.org/wiki/Simulated_annealing
Start your optimization process with a high "temperature" (leading to big changes), and let it "cool" over subsequent iterations.
http://en.wikipedia.org/wiki/Simulated_annealing
Start your optimization process with a high "temperature" (leading to big changes), and let it "cool" over subsequent iterations.
See http://visualwebsiteoptimizer.com/split-testing-blog/how-aqu... as an example of testing radical change.
See http://visualwebsiteoptimizer.com/split-testing-blog/using-a... as an example of subtle change.
In my opinion, it is entirely misguided to blame local maxima problem on A/B testing. The methodology doesn't dictate what is that you want to test. It is all about having a hypothesis and then collecting data to prove/disprove it. The post looks like it is blames poor results on instruments while wrong experimental design is the true culprit.
See http://visualwebsiteoptimizer.com/split-testing-blog/using-a... as an example of subtle change.
In my opinion, it is entirely misguided to blame local maxima problem on A/B testing. The methodology doesn't dictate what is that you want to test. It is all about having a hypothesis and then collecting data to prove/disprove it. The post looks like it is blames poor results on instruments while wrong experimental design is the true culprit.
Instead of just running A versus incremental-change A2, also run a B version that's radically different from A.
It's like Google testing 40-odd shades of blue to see which performs better, while failing to discover that the best color to use is (say) red.
It's like Google testing 40-odd shades of blue to see which performs better, while failing to discover that the best color to use is (say) red.
A few months or a year ago a Google designer quit when he was forced to split test one too many shades of blue. He described what a creativity sink the uniform fidelity to that sort of approach is. Of course, in a large organization there are always counter examples and counterpoints, but that particular Google designer was responding to frustrations inspired by recognition of this post's observations. Engineering has not yet reached the point where split testing is always (or usually, in my opinion), better and more efficient than the creativity of a trained expert.
Or that's what he claimed. For some reason, he forgot to mention that he went from a post-IPO behemoth to a pre-IPO company that was the hottest thing on earth at that point...
Here's the post: http://stopdesign.com/archive/2009/03/20/goodbye-google.html
I have a eight A/B tests currently running. Here's how long it took to launch each one. Spot the big bang test: ten minutes, fifteen minutes, fifteen minutes, 30 seconds, 15 seconds, an hour, three hours, three weeks.
Small A/B tests also have small risks associated with them. One of those eight touches a single word on my website. It is almost inconceivable that that one-word change could result in a POed customer sending me email. On the other hand, the big bang test can cause (and has caused) customer support issues for me, despite taking a great deal of time to minimize the impact.
In addition, big bang tests are often less conclusive than you want them to be. My current Big Bang test is significant at 95% right now, against the pivot I want to make. I am of two minds: on the one hand, I want to bow to that inevitability. On the other hand, I'm seriously wondering whether it is the pivot causing the disparity or if it is just implementation details of the pivot. In a standard A/B test, the change is the implementation detail. However, given that I had to adjust something like 30 files, I'm wondering if customers are really rejecting the pivot or whether they just think the graphic I made for it is exceptionally hideous. (I started an A/B test for just the graphic and it is, indeed, getting whupped versus what it replaced.)
Since the big bang A/B test frequently changes mountains and drops you at the unoptimized bottom of the new mountain, you're left either trying to do hillclimbing on two hills in parallel (which is OK, as long as you don't mind your engineering team cutting out your intestines and using them to strangle you) or doing the deeply unsatisfying "Well, I really feel better about B, so we're going to hop over there and then start hillclimbing, then pat ourselves on the back and assure ourselves it was the right decision all along."