Optimizing Web sites: Advances thanks to Machine Learning
Machine learning has revitalized a nearly dormant method, leading to a powerful approach for optimizing Web pages, finding the best of thousands of alternatives.
Figure 2: Relative values of variations in Web elements
An interactive simulation program can show the value of all possible combinations. This can be useful if you would like to see, for instance, how much you would lose in overall desirability using the boss’ favorite design instead of the best one. A very simple market simulator for communications appears in Figure 3. This one assesses the relative impact of print advertisements.
Figure 3: Screenshot of a simple communications-related simulator
The simulator allows you to change all features that were varied in any way—and to see the effectiveness of the resulting combination. Here results also are compared to a simple spot ad and the worst possible variant of a display ad. As the controls are changed, the display changes in real time.
Using real behavior on the Web, each visitor would see only one variant, and so you would need several thousand visitors. Each visitor would be randomly assigned to one of the test pages and behavior would be recorded. This is far fewer than the million or more reported with some A/B tests, and has the additional advantage of not exposing many people to a sub-optimal page.
If several thousand still would be too many, you could go to a survey-based approach. In this, a few hundred participants would be recruited and would see several alternatives from the full set. The survey would have the advantage of being able to select people with strong interest and/or use of the category, and of getting additional information about the individuals and their subjective evaluations.
Either test approach can provide far greater depth of information than the standard A/B method. And they will more than repay the extra time and effort.
There are many more details. You can find these in Artificial Intelligence Marketing and Predicting Consumer Choice.
 See, for instance, Wittink, D. (2001), “Forecasting with conjoint analysis,” in Principals of Forecasting, J. Scott Armstrong, ed., Boston: Kluwer Academic.
 Gelman, A. et al (2013) Bayesian Data Analysis (3rd edition), (Boca Raton, FL: Chapman & Hall) gives a through explanation of Bayesian methods and discusses hierarchical models
Bio: Dr. Steven Struhl is the author of Artificial Intelligence Marketing and Predicting Consumer Choice (2017, Kogan Page), Practical Text Analytics (2015, Kogan Page) and Market Segmentation (1992, American Marketing Association; revised 2013). He is founder and principal at Converge Analytic, specializing in advanced analytics and marketing sciences. He has over 30 years’ experience with a wide range of industries, government and non-profits. firstname.lastname@example.org