I'd be slightly more careful about making the claim you are making (though in your defense, nearly every person who makes a boastful post about their A/B testing is makes the same mistake).
For starters, you are picking an incredibly noisy metric, taking the maximum observed conversion and the minimum observed conversion and then dividing them. To better illustrate this, I wrote a little test script to simulate what you did many times. It assumes you have 4 email variations, send 125 emails each, and the average conversion rate is 5% across the buckets. Then it takes the minimum converting bucket, the maximum converting bucket, and divides them. Then does that 10k times and averages out what the gain would be by this metric, which is ~2.7x (code here: https://gist.github.com/3846278). 2.7x is a lot, but of course in this example it is pure noise, every single bucket converts exactly the same over the long term.
This metric also is particularly unhelpful because your goal is to have one high converting email. The metric you have chosen will be artificially boosted if you happen to have one horribly converting email. While that is potentially interesting fodder in just how differently various emails can convert, having one horrible email doesn't help you very much.
A better way for you to do this analysis is to dump your raw results into a mathematically sound A/B test calculator (I'm a bit of a homer but I don't think you can beat ABBA [http://www.thumbtack.com/labs/abba/]), then look at the confidence conversion ranges of the various emails and only make claims based on that. Like... I tried 4 versions of email copy and got my conversion rate up to X% (+/- some hopefully small confidence interval)! One of the emails was a real stinker and only converted at Y% (+/- confidence interval), thankfully I A/B tested first and didn't end up getting stuck with that one!
I think one thing to be careful about is not to read too much into response from your friends... they are already biased. That said, the difference within that set is interesting.
That makes sense...that was why we wanted to point out relative performance rather than absolute numbers. It is also probably fair to point out that the content and tone of our messaging differed quite a bit, which explains some of the differences.
When does family/friends announcements turn to spam? I know many marketing email providers are weary of writing to your entire address book of people who did not opt in.
We actually looked at using MailChimp for this, but we didn't want to go outside of the TOS. So, we ended up merging the mail locally using Thunderbird and sending it from our personal accounts. I am sure there is more to be said about the line between the two, however.
From the graph labels, it appears the subject line which moved the most relevant info to the front won, as opposed to leaving out the startup name or pushing it to the end.
But, they don't mention the sample size, and they do mention that there were other changes in the email content and call-to-action. Nor do they make any mention of statistical significance.
So we can't tell what actually helped, only that perhaps (if we assume did their significance-testing correctly) that some set of changes can make a big difference.
Pasting in the emails was going to make the post very long, but to summarize, we presented Meritful as a tool for students to build a positive web presence, as a tool for teachers and mentors to run projects and give feedback to students, as a tool to help with college admissions by creating an impressive portfolio and interaction from respected folk and another comparing Meritful to existing products for teens.
The sample size was roughly evenly split between the total number of contacts we had.
For starters, you are picking an incredibly noisy metric, taking the maximum observed conversion and the minimum observed conversion and then dividing them. To better illustrate this, I wrote a little test script to simulate what you did many times. It assumes you have 4 email variations, send 125 emails each, and the average conversion rate is 5% across the buckets. Then it takes the minimum converting bucket, the maximum converting bucket, and divides them. Then does that 10k times and averages out what the gain would be by this metric, which is ~2.7x (code here: https://gist.github.com/3846278). 2.7x is a lot, but of course in this example it is pure noise, every single bucket converts exactly the same over the long term.
This metric also is particularly unhelpful because your goal is to have one high converting email. The metric you have chosen will be artificially boosted if you happen to have one horribly converting email. While that is potentially interesting fodder in just how differently various emails can convert, having one horrible email doesn't help you very much.
A better way for you to do this analysis is to dump your raw results into a mathematically sound A/B test calculator (I'm a bit of a homer but I don't think you can beat ABBA [http://www.thumbtack.com/labs/abba/]), then look at the confidence conversion ranges of the various emails and only make claims based on that. Like... I tried 4 versions of email copy and got my conversion rate up to X% (+/- some hopefully small confidence interval)! One of the emails was a real stinker and only converted at Y% (+/- confidence interval), thankfully I A/B tested first and didn't end up getting stuck with that one!