Omniture Articles

The Merchant Versus The Machine

In the waning days of the bubble, the good folks at landsend.com held a simple challenge: three technology vendors were invited to pit their best algorithms against a single online merchant.

May the best man (or machine) win.

The vendors were best in class, including ATG and two others. And each set up their system with real products and real data. And I assure you from direct experience, everyone played to win.

Needless to say the machines lost. Round one to the merchant.

Now optimization is back. And so is the question. Are merchants better at suggesting what shirt goes with what pants, or what case goes with what camera, or is a highly advanced algorithm? Can sites like Pandora.com, which recommends music, outperform a DJ? Can Amazon get us to buy through their relentless product associations?

Seems, we may as well be asking whether John Henry is going to dig faster and farther than a steam shovel.

My belief? The merchant will likely always win the head-on-head competition like the one at Lands End. And the debate is really a false one.

So what should you do? The potential of an optimization engine could revolutionize the way you merchandise. Or is the engine just a cheat that can never quite replace solid marketing experience and intuition?

The machine doing automated merchandising (and there are many flavors of testing and optimization approaches) can be far faster, handle a larger assortment, and be much more sensitive and responsive to changes in buyer behavior than a merchant. The most sophisticated approaches can factor in a volume or range of data that is impractical for a merchant to do with any efficiency.

But, in my direct experience, the machine is only as good as the merchant that is driving it. There is no true split - the merchant and the machine must become partners.

In that spirit, here are three areas where a good online merchant can use automated product suggestion to enhance their merchandising strategy:

Tip #1. Different tactics for different locations

Testing and optimization can be used anywhere that you normally make merchandising decisions: search results, cross-selling suggestions, impulse buys, landing page and home page product suggestions, even on the category page itself.

Home pages are a great place for merchandise optimization. Select a small set of products and have the algorithm continuously optimize the location and priority of the products. Track user behavior and target specific, merchant-chosen products to user segments based on on-site and prior visit category and product affinity.

Product pages can be extremely difficult. Accessories and other standard approaches can work. “Kitting” and bundling can enable a merchant to increase basket size. Our recommendation is to let the algorithm establish a basic set of product suggestions for each product, then have the merchant improve the suggestions manually for key products or categories.

Step #2. Consider the criteria on which the optimization engine will work

We implemented optimized product recommendations a while back for a merchant on a few category pages. We had a merchant pick a set of 12 products and set the system to display the best performing six in a standard product grid.

Our first attempt was a crushing defeat, performing more poorly than the existing product list.

But we then adjusted the criteria. Instead of ranking products according to what led to purchases of the products displayed, we simultaneously tested a version that optimized for which products were viewed more often and a version that optimized just for total order size.

Both of these alternatives beat the control group and our first attempt. And the winner increased revenue per visit by about 8%. Understanding how to make the algorithm responsive was a human activity.

Other winning approaches for clients include tuning the algorithms for the price point of product suggestions and even changing how quickly the algorithm re-evaluated the performance of the products. It matters.

Step #3. Narrow your options

The third tactic is for a merchant to pre-screen the alternatives for product suggestion, then let the machine refine them. You know your products best, so sometimes it makes sense to narrow the suggested product field to those that you know, from experience, will do well.

This is especially important for branded goods and for merchant-driven categories such as gift finders.

The role of the merchant here is powerful -- making decisions and effectively “taking control” from the machine. But the merchant can also take advantage of the unique power of automation to become substantially more productive and responsive.

With the right approach, you can even have two merchants compete. Each picks a set of products and sets the levers on the algorithm. The merchandising system routes traffic between the two and can give a concrete winner.

Quickie Case Study

Here's a short case study to illustrate what I'm talking about. In working with an online retailer of significant size, we were able to help them potentially increase annual revenue by millions.

On the category page of the category they hoped to test, the retailer showed visitors a six-product grid of best sellers. Working with them, we decided to test those products to see if we could significantly improve revenue per visitor (RPV).

The default was the six product grid which merchandized the items in the way the retailer always merchandized them: that is, randomly rotating the 13 best sellers in that category so that each got equal billing.

  • Recipe #1. For the first recipe that we tested, the retailer used the same 13 products and the same six product grid. Rather than being sorted randomly, however, they were sorted in groups by products purchased (in other words, the product group of six that had the most purchases which contained one of the products from the grid), based on one day's data.
  • Recipe #2. The second recipe again used the same 13 products and the same six product grid. It was again sorted by products purchased, but for this recipe it was based on one week's data.
  • Recipe #3. Again, same 13 products, same six product grid. This recipe sorted the products based on orders placed (whether or not the order included one of the products from the six product grid), and was based on one week's data.
  • Recipe #4. Same products, same grid, this time sorted by most-viewed products within the last week.

The third recipe -- which sorted the products based on the most orders placed within the last week -- won, hands down. It resulted in an over 8 percent increase in revenue per visitor, a matter of more than $1 million over the course of the year.

Better yet, the best predicted recipe, which would combine the most useful elements of all the recipes, could result in an increase of nearly $4 million annually.

In each of the recipes, the automated product suggestion algorithm created the suggestions, read the data, and responded. But the algorithm was far smarter with the merchant than without.

Without the merchant, the basic set of products would have been unreliable. Without the tuning, the algorithms had only a random chance of winning.

Working together, the merchant and the machine made money.