Supercharge Your Cross-Sell/Upsell Strategies with Graph Learning

Ivaylo Bahtchevanov

Importance of AI-Driven Cross-Selling and Upselling

Using AI to enable efficient cross-selling and upselling has become increasingly important in helping businesses increase revenue and profitability while simultaneously improving customer engagement and loyalty. In this blog, we’ll discuss the importance and benefits of using predictive analytics to power these motions and talk through some of the best approaches for optimizing these strategies in both B2B and B2C businesses.

Cross-sell strategies attempt to identify complementary products or services that existing customers are likely to add to existing purchases. Examples of these strategies include:

• Ecommerce platforms recommending additional products to add to online shopping cart (ie user checking out a phone on Amazon will see suggestion to add a phone case and headphones to the cart, or a user buying a grill on Home Depot site will see a dropdown to add a grill cover and spatula)
• A bank offering a credit card or other financial services to a customer actively opening a checking account
• Food delivery or restaurant platforms showing combos that pair with an existing order
• Sales team positioning discounted add-on products in a B2B enterprise deal

Upsell strategies encourage users to upgrade or move to more premium or higher tier offerings. Some examples include: 

• Provide upgrades or premium membership opportunities to users who might be interested in new features, more usage, or whose profiles are similar to other users who have upgraded
• Apps and gaming services offer upgraded subscriptions to base-level users who would benefit from a higher tier subscription
• Hospitality and travel services can upgrade customer bookings for hotels, flights, experiences, or other add-on services

To have a successful cross-sell/upsell strategy at scale, AI-driven predictions are incredibly valuable for a number of reasons. For one, there is no historical data that perfectly describes what the purchase recommendation should be since, by definition, the problem you’re solving is to identify offerings to combine that have not been purchased yet. What’s more, applying a blanket strategy or rules-based approach for generic recommendations across users will always result in mass-spamming, which in turn decreases engagement and eventually results in churn. Each recommendation needs to be very targeted and personalized in order to be effective or even well-received. Performing manual curation wouldn’t work here either since that approach doesn’t work at scale – since it’s impossible to capture all relevant purchase affinities, this results in missing most important opportunities for making a sale.

Algorithms that predict cross-sell/upsell opportunities are part of a broader set of tools designed to provide a highly personalized and improved customer experience based on proactively identifying a user’s preferences and purchase affinities. These predictions can have many downstream benefits for the business including higher user engagement and revenue per user, more effective marketing and outreach campaigns, increased sales team efficiency, and more.

Effectiveness of Graph Neural Networks on Cross-Sell and Upselling 

Graph Neural Networks (GNNs) have emerged as the leading machine learning approach designed specifically for processing data that can be represented as graphs. GNNs are able to capture complex relationships and dependencies between different entities in a graph and can operate on large amounts of heterogeneous data. Read more about the benefits and effectiveness of GGNs in our blog on GNNs.

In the context of cross-selling and up-selling opportunities, graphs can be used to represent relationships between different products, between customers and products, or between customers. GNNs then leverage information from the entire graph to identify new patterns and connections, which are leveraged to produce highly accurate predictions on what additional products or services users are likely to buy. 

GNN embeddings leverage signals from the entire graph, which captures similarities between different users/products, as well as signals from different types of interactions – including views, clicks, searches, purchases, notification opens, and more. GNNs are effective at capturing the diverse relationships between users, products, content, ads, and any other entity in the graph.

Figure: GNN captures co-purchase affinity by leveraging signal from co-views as well as purchase-after-view trends

More specifically, predicting affinities between users and new products/offerings that don’t exist is perfectly suited for link prediction – a GNN prediction technique designed to identify new links between entities. This method has proven to be highly effective in practice.

Amazon’s co-purchasing predictions saw improvements of between 30% to 160% over state-of-the-art baselines when incorporating GNNs (measured by hit rate and mean reciprocal rank), and improved their overall accuracy of recommendations by another 7%.

Figure: Amazon co-purchase recommender using GNNs outperforms existing benchmarks

The Kumo Approach

Kumo’s approach enables you to go from raw data to large scale GNN-powered predictions in minimal time-to-value. The user connects their tables directly, and Kumo will build your enterprise graph from the raw data, learning the relationships and interactions across all entities in your tables. Kumo provides a flexible no-code interface for quickly generating many concurrent predictions for any number of use cases. This powerful abstraction, called the Predictive Query, allows users to perform a comprehensive analysis on their users that then determines the optimal recommendation strategy for each user based on a desired outcome. 

With Kumo, you can build a comprehensive personalization strategy in which you predict the product affinities across all users for all possible products and offerings, and then identify the strategies most likely to succeed. A Kumo user can test out many hypotheses quickly by spinning up predictions that test the outcome of each strategy, and then identify the best set of actions to maximize cross-sell and upsell.

Taking this approach one step further, the user can then predict how these specific affinities will change in the presence of a GTM intervention. 

Some examples of queries you can run:

Identify best co-purchase recommendations for a set of users given you perform a sales outreach

Determine the impact of sending customer a discount coupon – and test the effectiveness across a range of times and templates

Determine the impact of sending a push notification – test across different mediums like email, mobile, in-app

Send a marketing campaign – run through a variety of scenarios and ads and identify effectiveness of each

Running a series of queries allows you to perform a comprehensive what-if analysis on new strategies and then to iterate quickly to maximize engagement, retention, and revenue.

Interested in learning more? Reach out to us here to try it for free!