An interesting nuance of traditional AI models is that they tend to look at the correlation between products considered and products purchased. In other words, people who have looked at soccer cleats have tended to buy shin guards, considering behavior on site or in store. However, there's a lot of evidence that suggests that when consumers are presented with these SKU driven ads in multi-choice ad formats, they start a journey that isn’t always linear.
As consumers engage with these ads, either through a view, a mouseover, or a click, what they buy is frequently not the SKU of original intent. Often, it's not even remotely close. Even in the best cases, a third of purchases made have a strong affinity to the original SKU.
It’s likely that the role of an ad like this is to simply bring the consumer into the store, whether it be virtually or physically. That certainly plays into the age-old principles of marketing, as well as our own experience as shoppers. Retailers know that once we’re in and have that first item in the cart, others will follow. While the role of the site is to monetize the available inventory, the role of the creative may best be seen as starting, rather than confirming, the final purchase.
This leads to an open ended question when setting up the neural network to evaluate recommendations. Does it make sense to evaluate consumer behavior on site, and ignore the behavior in ad?
Instead, a more valuable model may be to evaluate the behavior in both. For example, people who have seen a SKU advertised, or have clicked on it, tended to purchase it. The value of the SKU could be a better signal in the end. Even if the customer never purchases the specific SKU, it led to better outcomes down the road.
This won’t come as a surprise to many retailers. Advertising loss leaders are a time-honored way to get people into the store and put that all-important first item in the cart. As more and more consumers mix offline and online shopping, the data seems to indicate that behavior works with digital creative as well. This is also where many savvy retailers work to increase cart size with discount shipping at higher purchase levels. This is, of course, a great point to test in your advertising creative.
In conclusion, the best recommendation engine is going to offer creative SKU recommendations that increase the likelihood of the consumer becoming a customer. It’s also going to drive overall purchases, rather than simply display the best sellers. A more nuanced look will let the data drive, and train it to make product recommendations that results in a purchase directly from the ad, rather than just on the site.