Good post. I just want to expand a bit on what I think is an important issue, so forgive me for its length.

I think the image of the maze is the key to thinking about optimizing digital applications. In order to optimally solve a maze one needs to find the shortest sequence of decisions that will lead to the goal for each of the possible locations within the maze.

So, for example, you might have four possible decisions at each location in the maze (up, down, left, right). You need to figure out at each location which is the best direction.

The challenge is that

1) the reward for getting out of the maze only occurs at the final step and;
2) the value of a decision in one location is dependent on all of the future decisions.

Your digital application is in some sense like the maze. There can be delayed rewards - the conversion events might be several steps away from a landing page and as the post highlights there are multiple decision points where the optimal decision at one depends on decisions made at the others.

Unlike the maze, the digital app:

1) has different types of users who all might respond differently to each decision;
2) even if there was only one type of user, each decision taken only affects the probability that the user will respond in one way or another;
3) users can 'teleport' into the application at almost any location (via bookmarks etc.) and also exit from anywhere;
4) there can be multiple goals; and
5) each of the goals may have different values.

The optimal solution for the digital application is to globally maximize the value of the application by finding the optimal decisions, for each user, in each of the relevant locations, in the shortest amount of time (if factoring in the opportunity costs for presenting suboptimal decisions). One more problem, on the web information is often perishable, so whatever solution you might have converged to may not hold next year, month etc.

How to do all of that?

Personally, I don't think that many of the current mainstream optimization solutions will get you there. For one, most of the current solutions implicitly assume that your application is flat - so no real way to incorporate the user/application dynamics into the solution. Secondly it is unclear, to me at least, if hypothesis testing is the best approach for solving real-time dynamic optimization problems.

Of course it will depend on what you are trying to do.

If what you really want to do is more of a clinical trial than optimization -extract some learnings from a sample in order to generalize or use in an offline setting - then hypothesis testing is great.

If on the other hand you have a ‘Deep Blue’ problem, so learn to play the optimal game of ‘chess’ against different players, then you will probably need to use other methods.

Matt,

Thank you for your thoughtful reply. If you'd ever like to be a guest contributor here please let us know.

You identified the main challenge perfectly: "The optimal solution for the digital application is to globally maximize the value of the application by finding the optimal decisions, for each user, in each of the relevant locations, in the shortest amount of time"

To this day, you still won't hear very many marketers saying things like "this redesign is all about enabling each user to get what they want as quickly as possible!"

We do have a bewildering array of technology choices to sort through, but that spells great opportunity for organizations that get it.

-MrOpti