Why De-averaging Helps Assure Best Recommendations

As marketers and retailers who are now decidedly data driven, we tend to think that by seeing lots of transactions, we can predict what the “average” consumer preference will be. But that could be an illusion that de-averaging (finding a range of  right answers) can address. Let me explain:

Take something as “simple” as product recommendations. This is a broadly established discipline leveraging largely the same collaborative filtering approach that Net Perceptions pioneered on sites like Amazon back in the mid to late 90’s.

It makes sense that where the same customers commonly buy two products, future customers might similarly buy both. So if jelly is typically bought with peanut butter, showing grape jelly to your next prospective peanut butter buyer makes sense.

But is the best answer on average always the best? What about people who like honey instead of jelly? Or maybe a different type of jelly? Apply this to the more complicated relationships within your own product assortment and you’ll quickly find a lot of your recommendations are disproportionately “best” for a singular cluster of customers rather than true for all of them.

While this should not be taken to the point of diminishing returns, a set of best answers vs. just one for all creates additional opportunities. This process is called de-averaging and helps assure that your recommendations will deliver lift because that single best answer just isn’t as relevant to all customers.

Here’s another example, promotions.

For a retailer, an ideal world would be one where the promotion offered perfectly balances the incremental sales it brings in against the cost of the promotion to achieve the highest possible contribution.  This is harder to pull off than most retailers would like, but it’s coming increasingly within reach.

But a better world would be one where there isn’t just one promotion offer for all comers, but a range based on who it is being offered to. Retailers could then offer the lowest possible promotion (in terms of its cost to them) needed to trigger a purchase.

This isn’t some consumer dystopia where flinty retailers abuse hard working citizens. Not least because some people would see a higher promo value even while others got a lower one. Indeed, some might get a promo that costs the retailer less (e.g., a free shipping offer instead of $ off), but was in fact valued more highly by the particular consumer than what would have otherwise be offered.

Point being that de-averageing promos could take costs out for the retailer while still allowing them to capture more sales than a one-size-fits-all promo would. And for the consumer, they get a deal they still want.

Here’s another: re-targeting

Are all people who leave your webpage the same? It’s more likely that even two people who looked at same product actually do not have the same value for the retailer. Some may be massively more valuable – and should be far more aggressively bid for in a reacquisition effort. Conversely, others will be much less valuable than average and should be bid down or avoided entirely.

As with promos, there is an opportunity via de-averageing to take costs out by reducing your spend on bad/lower value customers while simultaneously boosting overall sales by making sure you don’t let the most valuable consumers slip entirely away.

Direct marketers will read this and say “no kidding.” Direct mail and e-mail have long practiced this kind of de-averaging. It’s called it segmentation.

What’s different about the above examples that make it worth separating de-averaging from segmentation?

First, de-averaging is done in real-time. It’s not an asynchronous segmentation/execution model. It has to happen in milliseconds. Serving a recommendation, popping a promo, or tweaking a real-time bid has to happen as soon as you see the consumer or the opportunity disappears.

Second, it requires predictive analytics to make the decision on what to do. This isn’t classic CRM-style segmentation. The latter is still extremely valuable, but it was not built to do the kinds of de-averaging we are talking about here. Especially since in most cases the consumer you are applying the decision to does not even exist in your database (or if they do, you don’t know it).

Applying predictive analytics in real-time was not broadly practical until the arrival of both cloud computing (the computational horsepower would have been prohibitively expensive) and the big data tools and methodologies required to dive them. But with their arrival – and suppliers dedicated to harnessing them – it’s possible even if you are not Amazon.

In fact it’s happening right now with a rapidly growing number of retailers. De-averaging will continue to spread quickly – opportunities to take out costs AND grow revenues are too precious. And as more retailers take advantage of it, they will put pressure on the laggards to follow suit or suffer a competitive disadvantage.

Richard Harris is co-founder and CEO of Intent Media.