In my entire existence as a business I have never seen someone by X licenses plus Y [license plus CDs], and indeed even explaining the difference to my customers would be tricky. So I opted to make it impossible -- if you ask for 5 copies of the software via download you get five copies, if you then ask for the CD instead it will set you up for 5 copies and 5 CDs.
The cart actually remembers everything that you've put into it within the current page, even if you close the lightbox. If I wanted to save it there are options but a) dirty hack and b) most of my customers have no need for it.
Your sample size is also quite low (94% increase = from 8 conversions to 15), and you're using absolute values in the graphs. Did traffic to the page stay constant over all that time?
That last question appears to demonstrate a misconception about A/B tests. I did not test the old cart serially with the new cart -- I've done that sort of thing before, but the results are automatically suspect because factors other than the variable you're testing are constantly changing. An A/B test tests the old cart and the new cart at the same time -- when you open purchasing.htm Google flips a coin and cookies you up with the results. Heads you get the old cart. Tails you get the new cart. No matter how many times you go back you get the same cart (until I terminate the test, obviously).
This means that I'm able to have confidence in the results despite this week having traffic far above my typical values, due to Valentine's Day. (Certain holidays are almost always good to me. Why is outside the scope of this post.)
The sample size was not 8 or 15, incidentally. It was two groups of over a hundred (prospects, not customers). While I'd prefer groups of over a thousand for the obvious reason that it implies I'd sell ten times more software, in stats terms that doesn't make the experiment more valid, it would just decrease the size of the confidence intervals by roughly a factor of sqrt(10), and it might also increase the confidence in the significance test (that was the second 94% value, see the writeup).
The image clearly shows that the result is statistically insignificant. This is very important as you could be misguided. The general thumb of rule for running full factorial A/B test (like the ones GWO does) is that you require about a baseline level of million page views with about 5% conversion rate in order to get successful results in one week (including weekends).
My understanding of the word "full factorial" means that you have multiple design elements under test at the same time and you're testing all possible combinations. For example, you are testing two alternative images and two alternative headlines. This gives you 2 x 2 = 4 possibilities to show to any given user. As you increase the alternatives for each factor and increase the number of factors, the total number of alternatives grows in a combinatorially explosive fashion and you might indeed need a million page views and 5% conversion rates. (Heck, six factors with 6 options each and even with a million viewers you'd have less people seeing each combination than I did, unless you started pruning them early.)
But I'm still only testing two alternatives of one factor. I mean, yes, that is included in the definition of "full factorial" but it makes an absolute hash out of that rule of thumb. Two choices total means the stats test is simple and does what it says on the tin: 94% chance that new cart outperforms old cart, exact magnitude of outperformance bounded by calculable confidence intervals.
You can consider 94% insignificant or significant -- your call really. If you chose p = .05, its insignificant. If you choose p = .1, its significant. It costs me very little (except opportunity costs) to keep the experiment going but 94% is good enough for me personally to claim a win out of it.
Although in the image, nowhere could I see p value, I have verified it by myself that there is 95% significance. And I agree with you it depends on perspective on what you consider significance.
Sorry, should have been more specific: I meant that you use absolute $ values in your sales-by-month graph -- that must also be dependent on traffic. Obviously the A/B test is not.
In my entire existence as a business I have never seen someone by X licenses plus Y [license plus CDs], and indeed even explaining the difference to my customers would be tricky. So I opted to make it impossible -- if you ask for 5 copies of the software via download you get five copies, if you then ask for the CD instead it will set you up for 5 copies and 5 CDs.
The cart actually remembers everything that you've put into it within the current page, even if you close the lightbox. If I wanted to save it there are options but a) dirty hack and b) most of my customers have no need for it.
Your sample size is also quite low (94% increase = from 8 conversions to 15), and you're using absolute values in the graphs. Did traffic to the page stay constant over all that time?
That last question appears to demonstrate a misconception about A/B tests. I did not test the old cart serially with the new cart -- I've done that sort of thing before, but the results are automatically suspect because factors other than the variable you're testing are constantly changing. An A/B test tests the old cart and the new cart at the same time -- when you open purchasing.htm Google flips a coin and cookies you up with the results. Heads you get the old cart. Tails you get the new cart. No matter how many times you go back you get the same cart (until I terminate the test, obviously).
This means that I'm able to have confidence in the results despite this week having traffic far above my typical values, due to Valentine's Day. (Certain holidays are almost always good to me. Why is outside the scope of this post.)
The sample size was not 8 or 15, incidentally. It was two groups of over a hundred (prospects, not customers). While I'd prefer groups of over a thousand for the obvious reason that it implies I'd sell ten times more software, in stats terms that doesn't make the experiment more valid, it would just decrease the size of the confidence intervals by roughly a factor of sqrt(10), and it might also increase the confidence in the significance test (that was the second 94% value, see the writeup).