Choosing a good recommendations system can be confusing, given that there are so many companies offering these services. There are many considerations to identify the most suitable e-commerce product recommendation system, such as business size, complexity, business objectives, platform integration, and cost. In this section, we discuss six key questions you address in your selection process:
1. How do I identify a best-of-breed personalized product recommendation engine?
There are many simple product recommendation engines, and the concept of AI has many different definitions for many people. So, reading the tagline “AI-based recommendation system” can mean a lot of different things. So how do you distinguish the powerful AI-based systems from the wannabees?
First, let us talk about breadth. A good product recommendation engine should allow you to select from multiple algorithms, as opposed to relying on only a few algorithms. This is because, as discussed above, you want to seduce your visitor in different ways in different places on your e-commerce site and in different stages of his visitor journey. Another feature that increases breadth is the ability to combine rules with each algorithm. These combinations can result in a high number of very specific algorithms for very specific use cases.
Secondly, depth is important. This is about the power of the algorithms and the ability of these algorithms to improve their conversion power over time when more and more data is becoming available. The power of the algorithms is for example driving the ability to hyper-personalize. So, the different visitors get different recommendations, based on the profiles related to these individual visitors? Another way to determine the depth of these algorithms is to determine which AI concepts are applied. Best of breed algorithms apply deep-learning, machine learning, natural language processing, embeddings, and neural networks.
2. Is my e-commerce site big enough for recommendations to work well?
The key factors for personalized product recommendations to work well are visitor traffic and product assortment.
The more visitor the better, and the more visitors come back regularly the better. Returning visitors allows the system to create and maintain an insightful visitor profile that can drive the personalization of future recommendations.
The more products in the assortment and the more variety within the assortment the better.
As a rule of thumb, e-commerce sites selling at least 1,000 products with traffic of at least 20,000 unique visitors per month benefit really well from recommendations systems. Of course, this does not mean e-commerce sites with smaller numbers cannot benefit from recommendation systems, but as argued above, the more data there is, the better the recommendations become.
3. Is the system a black box, or can you control how the system behaves?
What we hear a lot from the e-commerce managers we speak with is that they do not like systems to act as a black box. To counter this, there is a couple of things to address.
Transparency of the algorithms, and the ability to configure these
Can the supplier be clear and transparent about how the algorithm works? The first step in getting control over the system is understanding how it works. This first step is where a lot of systems fail already.
Once the understanding of how the algorithms work is achieved, what influence can you have, as an e-commerce team, on the configuration of the algorithm? Can you influence the signals the algorithm considers? For example, the optimal configuration of a ranking algorithm can differ from one category to the other. In fashion, for example, you may want to see “newness” as an important factor to rank products for women, but for men “margin” may be more important. And for the category of items on sale “stock levels” maybe most important.
Manual insertion and associations of products
Sometimes e-commerce managers want to push certain products within certain recommendations. It is therefore usually important to have the option to manually insert chosen products within a widget to fulfil your business objectives. For example, a toys vendor may decide to sell the game “Zelda” with the newly released PS5 console, in the first 3 weeks after release. Or a retailer may “sell” the first two slots within a widget to a specific brand or supplier. A retailer may also manually insert specific products into a widget to get rid of some remaining inventory.
Steering information to assess the effectiveness of the widgets
Your system should come with a dashboard that tells you exactly how your widgets perform and why. It should provide insights that allow you to determine where the system needs to be tweaked. Which widgets placement should be adapted to get more views? Which algorithms should be adapted to incite more clicks?
4. Can you combine product recommendations with site search and email marketing?
Personalized product recommendations can be integrated into your search bar. It is possible to use search queries as an additional signal to make recommendations even more powerful. Recommendations can be made to appear on the search bar while the search query is keyed in. These personalized product recommendations are effective as they are related to the query. As a result, conversion increases significantly because visitors can find perfectly matched products directly from the search, rather than moving from search through several further pages before they find the ideal product.
Similarly, inserting personalized recommendations into your email campaigns allows you to hyper-personalized your campaigns and boost the conversion of these campaigns. Email campaign systems segment visitors’ data for campaigning, while hyper-personalized recommendations promote the products that the visitors really care about. This powerful combination assures maximization of conversion.
5. How easy is the integration?
Typically, integration addresses these four areas:
- Uploading the product catalog onto the recommendation system
- Uploading historic transactions to jumpstart the recommendation engine
- Tracking visitor behavior on the site
- Rendering recommendations to the visitors.
The product catalog needs to be updated very frequently to ensure a proper account of stock levels and prices. API-based integration is best for this, but a frequent sharing of a product feed is also possible.
Historic transactions can be integrated by sharing a simple feed, or through an API key.
Clickstream listening, enabling the tracking of visitor behavior on site is often done through a listener file, that is triggered by an inserted div tag within the e-commerce platform.
It’s important to ask potential system vendors how long the implementation will take, and if their system already has pre-built integration with your e-commerce platform. Implementation can take between a few hours and a few weeks, and it depends on factors like the e-commerce platform, the degree of pre-built integration, and the complexity of the product catalog.
6. Is this vendor going to delight me as a customer?
The interaction you will get with a vendor in the sales process is often a good predictor of the interaction you will get in the service delivery after you sign a contract. What is important to assess is whether the sales team of your vendor and the customer success team of your vendor is the same team or a completely different function. We hear stories from the market where contracts are signed and are barely turned over to operations or customer success, where operations do not even know what exactly was agreed in the sales process.
Better is to find a vendor where sales and customer success is part of the same team, where there is no paid service or consulting, but instead, customer success is quantified and agreed upon within the contract upfront.
Also important is to assess how important you are to your vendor. You need to open your sensors in the sales process to get a proper assessment of this. You need to feel that your vendor cares about your success and will do all in his power to make you succeed.