Science Behind AI Supercharged Conversions for An Online Retailer
Analysis of AI Supercharged short-term behavior data can quickly recognize shifting consumption patterns and help retailers boost conversions
The COVID-19 is impacting e-commerce in lasting ways. Over the last four months, the move to online shopping has accelerated and expanded into categories such as groceries and home improvement. Consumers under lockdown are spending more time online, researching, and shopping. While e-commerce revenues have surged overall, economic uncertainty is shifting demand to essentials, typically lower-margin products and categories.
Retailers have been investing in AI technology to better understand consumer behavior over the last few years. Machine learning algorithms, for example, have helped marketers match audiences with advertising content and digital commerce improve merchandising recommendations. But as we have learned from the global pandemic, the world can change swiftly.
Learning from past data is key to machine learning. Typically, data from many sources such as clickstream activity, profile and transactional data is aggregated over long periods of time. More the data, better the prediction results. However, using data over longer periods reduces the ability to respond to short-term changes. Today we need a faster reaction time, as well as highly accurate predictive data analysis, to meet the swift changes of the new normal.
Short-term behaviors are the new normal for Online Retailers
Looking back on COVID-19, the early days of the pandemic saw a precipitous drop in consumer spending during the month of March. This was followed by a spike in April, which was surprising from the point of view of its timing as well as the amount of consumer spending. The benefits of economic stimulus and gains in the stock market seemed to buoy consumer confidence and offset some of the continued anxiety about health, economy, and jobs.
Four months after the pandemic started, it appears that significant gains in categories such as groceries, home improvement, kitchen, and furniture will continue well after the pandemic ends.
However, as swings in the stock market fluctuate week-to-week, consumer behavior also seems to shift continuously. This new norm for online retailers is likely to continue for some time to come as new outbreaks in the pandemic ebb and flow. With online conversions at only 3% in the pre-COVID-19 days, converting casual browsers to buyers was already an uphill task for online retailers. Changing consumer behavior patterns is further adding to that complexity, even as operating costs rise and margins erode.
Speed to sense is the new AI superpower
Machine learning algorithms use the abundance of historical data to memorize repeating patterns. Shifts in consumer behavior – especially sudden changes due to external economic and social events – presents a significant challenge to these techniques. New methods of machine learning must not only recognize these shifts in short-term consumer behavior but do so quickly with limited data.
ZineOne has developed a new way to model short-term behaviors from using only clickstream data – which contains customer activity on a site – as an ordered sequence, which is termed as Customer DNATM. As the encoding and analysis of human DNA revolutionized scientific insight, their patented algorithms have sequenced spatial and temporal consumer traits, fundamentally changing our ability to anticipate behavior. These algorithms successfully boost learnings from short-term behaviors and dampen or diminish those from data seen over a longer time period. This “speed to sense” is a new AI superpower.
An application of this technology is to identify visitors who plan to transact from just the first few clicks of their site visit. Or identify a dissatisfied customer just as they show the first sign of leaving. A Top 10 US-based department store leveraged this technology to analyze short-term behaviors on their web and mobile web e-commerce sites during COVID-19. The business team wanted to identify on-the-fence shoppers – site visitors who are less than half as likely to buy as the sitewide average – while they are still on the site. And give a real-time, personalized coupon to these on-the-fence shoppers to nudge them to purchase within their visit.
The results were very encouraging – these new algorithms were not only able to successfully identify fence-sitters with more than 90% accuracy, they also saw in-session, real-time offers increase conversions up to 50%! All this was achieved by limiting offers to non-buyers, which minimized “offer cannibalization” and maximized the effectiveness of the “offer budget.” Application of these novel machine learning techniques to short-term behaviors has also produced significant results for other online retailers.
Let’s understand some key challenges that were overcome to develop “speed to sense”:
Focus on in-session behavior analysis
A deep and comprehensive analysis of each visitor’s in-session behavior is essential in order to accurately understand their intent and readiness to purchase within that visit. For example, distinguishing prospective buyers from fence-sitters requires understanding how directed a site visitor is in their search and browse behavior. The analysis is particularly challenging as every activity of a visitor must be processed in real-time as they navigate the site, which requires significant processing capacity.
Discern varying customer behavior across visits
Continuation of behaviors across visits is important especially if the previous was recent or frequent. However, behaviors can also vary significantly across multiple visits, based on varying needs and motivations such as shopping for a product that needs urgent replacement versus a discretionary use product, shopping for self or a gift for an occasion, etc.
Perform early predictions
This point is crucial as it’s important to make predictions soon after a visitor arrives on your site, within the first few clicks. This ensures a high reach and ability to show on-the-fence shoppers persuasive actions. However, early prediction of purchase propensity is not trivial.
Take instant action on early prediction with AI supercharged insights
The term “instant” is key here as near-term predictions can change quickly. Consumer behavior changes rapidly, so purchase predictions for a session must be used immediately to trigger actions, and need to be computed for every subsequent visit separately.
AI that delivers speed-to-sense can enable organizations to adapt their e-commerce offerings on a continuous basis while personalizing consumer experiences and communications to what is happening on the site in real-time.
What does an AI supercharged ssuperpower look like?
In today’s crisis-ridden environment, a discount mindset is expected to set in, especially for non-essential products. This will make it even more challenging for businesses to not only to cut through the noise of innumerable offers and stand out to their site visitors but also to maintain margins. To further understand in-session personalization, let’s consider the following customer scenario.
1 | Customer visits the online storefront. Jason is browsing a retailer’s site for a new pair of running shoes. He’s looked through their products a few times before in the previous weeksbut hasn’t added his choice to his online cart.
As soon as Jason lands on the retailer’s website again, his real-time presence is recognized and his long-term activity data is also brought into consideration. The conclusion is that he is still looking for the right pair of shoes.
2 | Customer browses, but unlikely to check out. Jason thinks he’s found the right fit—a pair of lightweight, long-distance running shoes. He adds the shoes to his cart, where he is able to see the final price. A bit put off by the $85 total, he switches back to the products page and continues browsing.
Through a purchase propensity ML model, the business concludes that Jason’s purchase propensity for today’s browsing session is low – he has been viewing shoe options for weeks without a purchase. He did, however, place a pair of shoes in his cart, putting him in an actionably receptive state for a personalized offer – Jason is an on-the-fence shopper at this point.
3 | An optimized offer is delivered to the customer. Before Jason can leave the site without transacting once again, a personalized message appears on his screen: “Jason: Order in the next 10 minutes and receive $10 off your order!” The countdown begins.
The site is running a special offer: free shipping on orders of $75 or more. From past transactions, the business knows that Jason is price sensitive and he typically waits for a good deal along with free shipping. A margin analysis based on the real-time value of Jason’s cart and his previous transactions is performed, and a potential offer of $12 off is calculated — a personalized offer that can influence Jason’s likelihood to buy while still maximizing the store’s profit margins. However, applying a $12 deal to Jason would still not qualify him for free shipping. As a result, the $12 calculation is reduced to an optimized $10 offer. The countdown clock adds a layer of urgency to the conveniently discounted price and free shipping.
4 | Checkout is successfully completed. Satisfied with the $10 discount and urged on by the 10-minute limit, Jason uses the unique code in the pop-up to complete the purchase.
The offer presented to Jason can be redeemed only once, and only within the next 10 minutes. This allows Jason to successfully check out at a price point that delights him, gives him free shipping, and maintains profit margins for the retailer.
Analysis of the AI superpower
Providing every visitor with an offer is not the silver bullet to increasing conversion. As always, targeting the right visitor with the right offer is critical for success. By analyzing short-term behaviors and predicting likely in-session outcomes in the near term, retailers can answer these key questions:
- What is each site visitor’s purchase behavior?
- Is the visitor someone who will not buy under normal circumstances but can be influenced to buy in the same visit? In other words, they are the on-the-fence for that purchase.
- Is the visitor likely to buy without any incentive? If so, then leave them alone.
- Will the visitor not buy even with an incentive?
Clearly, the best option here would be to target the fence-sitters to increase conversions and secure immediate business gains. Once the business has identified the on-the-fence visitor, the next step is to determine an action that will influence them to buy in the same visit. Here the questions to consider are:
- What’s the discount amount most likely to influence a transaction for this particular customer?
- What’s the value of the merchandise of interest or in the visitor’s customer’s cart?
The quest to study shopper behavior and make purchase predictions is not new. Applying in-session machine learning algorithms can provide retailers new superpowers to personalize pricing and promotions. Instead of maximizing reach by issuing offers indiscreetly, retailers can drive incremental revenues by converting the largest number of influenceable visitors and maximizing the impact of an offer budget. Such a 1:1 customer engagement works during shifting consumer behavior and delivers personalization well beyond a rudimentary level.