As I’ve reflected on the wealth of data available to certain retailers and advertisers, I’ve generally assumed that it was beneficial to have a granular knowledge of customers, particularly with respect to their purchases, habits, demographics, psychographics, social relationships, and the like. I tend to have some faith in the power of predictive modeling, so with a large enough data set one ought to be able to have some confidence in predicting that, if a customer with certain characteristics buys X, she might also be interested in buying Y as well. The cost of making a ‘pitch’ for Y is virtually zero for an online retailer – a bit of screen real estate, perhaps a bit of nuisance for the customer, plus some coding and maintenance that theoretically amortizes over hundreds of thousands of customer interactions. It seems, then, that it shouldn’t be hard to come up with a business case for this strategy if one has any confidence in the forecasts one generates.
The humble grocery store, in my view, has a phenomenal implementation of this concept. I’ll assert that most people shop at no more than a couple of grocery stores, and so each grocery store gets a fairly comprehensive view of a person’s typical grocery list thanks to those ever-present loyalty cards. At the point of sale, the store can deliver coupons that respond directly to the items in the customer’s basket. Suppose the customer just bought a bottle of Heinz ketchup. Maybe he will get a coupon for fifty cents off his next bottle, to keep him in the family? Maybe he will get a coupon for fifty cents off his next bottle of Hunt’s ketchup, to encourage him to switch? Or perhaps the coupon will be for hamburgers or hot dogs, which go nicely with ketchup? Essentially, the grocery store could probably deliver meaningful coupons (i.e., cross-sales) without any historical information about the consumer – but perhaps building this store of knowledge over time allows them to refine their methods. I don’t know if this actually happens, but I wouldn’t be surprised if customers who frequently purchase expensive brands are pitched different items than those who frequently purchase store brands.
However, I think this trick is harder for an internet retailer to pull off. Since I will be projecting from my own habits, I should describe some of them for context. I rarely window-shop online, although I will comparison-shop if I’m already pretty confident that I want to buy, say, a new laptop. I’m not particularly loyal to specific retailers, although for particular categories of ‘need’ my defaults tend to be pretty stable (e.g., Amazon.com for books and, increasingly, digital music that’s remarkably cheaper than iTunes; Drugstore.com for certain essentials that are hard to find in my neighborhood; Bodybuilding.com for well-priced vitamins and supplements). So no single retailer has anything close to a comprehensive view of my shopping habits (in contrast, say, to my grocery store).
If my habits are close to those of a typical shopper at an internet retailer, it is easy to see that they will struggle to deliver worthwhile recommendations. My experience with Amazon, for example, bears this out. My purchase history of late has largely consisted of video RPGs for the Playstation 3, and non-fiction generally pertaining to business and finance. Amazon doesn’t know much about my preferences outside of these domains, so it will tend to have the most confidence at recommending me (a) other video RPGs, although bizarrely it has not yet ‘learned’ which systems I own; or (b) other non-fiction generally pertaining to business and finance. Fundamentally, this is no more sophisticated than organizing products by genre in a bricks-and-mortar store. What is more, Amazon is likely to conclude that I’m most likely to buy other popular items in those genres, which will lead them to recommend products that I’m already likely to have, or likely to know about, thereby reducing the extent to which their granular insight about me is incrementally useful.
Similarly, if Amazon were to look for correlations (e.g., customers who bought X usually bought Y at the same time), its most likely conclusions would also be (and have been, in my experience) fairly trivial – for example, a strategy guide for the video game in my cart, or a set of practice problems in the field where I’ve just bought a textbook. Of course there can be value to presenting cross-sell options at the right time, but it does not take proprietary insight to deploy the strategy I’ve described. Unless, for example, it’s more efficiently scalable to apply an algorithmic approach rather than identifying a priori which items are ‘accessories’ for other items (I am skeptical of this, though).
A question I’m not considering here is whether attempting to cross-sell, as an internet retailer, may actually be detrimental, thanks to the paradox of choice (although, see the link and comments at one of my favorite blogs for a more rigorous and skeptical discussion), which posits (in brief) that consumers can actually become overwhelmed and less likely to make any decision (i.e., purchase) whatsoever as more options are presented to him. If Amazon were most likely to recommend other finance textbooks as I’m considering the purchase of one particular title, could that actually lead me to buy none at all? I’m not sure about this, but I’m sure retailers can fairly easily test and optimize the number and type of recommendations to make.
As a consumer, I’d find it amusing if an Amazon were able to recommend me music or clothing on the basis of my taste in video games, for example; but I suspect they would struggle to make a good recommendation, and I’d assume their most successful recommendations would again be products that are likely to already be popular, which reduces the proprietary ‘edge’ of such a strategy (relative to, say, naively recommending the most popular product in a vertical).
I still believe fundamentally that granular consumer data ought to be useful for generating non-trivial cross-sales, but I’ve become less confident in its potential the more I’ve thought about it.