Archive for the ‘recommendation engine’ Category
Some of the online retailers we talk to, realize the value of cross-sells and up-sells and decide to embark on a “DIY path”. I may be biased here but I think product recommendations are one area in which you don’t want to reinvent the wheel.
There are 2 reasons why a DIY approach does not make sense in this case. The first is the development cost of building a system that works, and the second is the learning curve of optimizing it.
So how difficult is it to develop a recommendation engine? online retailers who decide to build a recommendation engine are not aware of the various components that need to be in place. Here are just a few them:
- The recommendation engine should track every major activity shoppers perform on the site including viewed products, categories and brands; items added to the shopping cart and purchased ; search keywords they used ; traffic source visitors arrived from, Geo-location data, and the list goes on…
- The system must support multiple recommendation types and should be able to display the right one (and more than one on a single page) based on where the user is at the purchase funnel (if you have less than 10 algorithms your system is extremely naive)
- Finding correlations between items/users is easy. The hard part is to choose which correlations should be taken into account and which should be ignored
- The system should have built in a/b testing and reporting capabilities so that it could be optimized and demonstrate its value. This point is very critical as few online retailers actually measure the impact of their homegrown systems
- The system should have an interface that allows marketers to control the outputs of the recommendation engine based on different variables.
The second reason I mentioned is the experience it takes to optimize such a system. There are many things that will determine the impact a recommendation engine will have on the business, and if it’s your first time building one and it’s a one-off project there is no chance you are going to know them or invest the time to learn them. In fact you are probably going to develop a very naive system and you’ll stay with v1 for a long time. You will not test different widget designs or various placements on the page.
The decision to build or buy a recommendation engine should be ROI based. You need to consider the impact it will have on your business versus the cost of development. Building a very naive system may seem cheap but will also likely to deliver poor results and without supporting systems you’ll never know it.