The Uncanny Valley of Targeted Marketing

Scott Howe of Acxiom spoke at Harvard last week as part of the Topics in Privacy series. I’ve been really interested to follow the steps Acxiom is taking to set an example in the advertising industry to engage more directly with consumers through Howe talked at length about the philosophy behind the site, its success in the first couple months, and addressed some of the early criticisms. I’m encouraged by what Acxiom is doing, but I walked away from the talk with more questions than answers.

Howe shared some statistics on the site after it first opened in September—they had 500K visitors in the first month, and only 2% of visitors have opted out as a result of logging in (I have to wonder, if that’s a[n intentional] design flaw, and that the option is buried or hard to find). He also shared that 11% of people made corrections to the data, most often addressing political party, income, education, marital status, and occupation. Howe also shared that the site has very low return rates so far, i.e. that once logged in, people aren’t coming back. He acknowledges that individuals won’t have reason to come back until the value of updating and maintaining a relationship with Acxiom is more clear to consumers.

Acxiom, like other data brokers,  is in the business of collecting, cleaning, analyzing, and segmenting consumer data from all kinds of sources. Aboutthedata only serves to show the demographic data, and some of the inferred insights about demographic data perhaps based on behaviorally tracked data. For example in my profile I am “inferred married” and it gives me the option to declare that I’m married. But Acxiom doesn’t expose the 70+ proprietary market segmentations* it has developed to describe me to marketers. I don’t get to object to being called a “Rolling Stone” or a “Midtown Minivanner” because these segmentations or “clusters” are fixed based on demographic details like household age, marital status, income, and “urbanicity.” Acxiom doesn’t expose this life stage segmentation to you, but I imagine mine is likely incorrect, given that Acxiom thought I owned a truck (I own no vehicle, let alone a truck).


But aside from these broad generalizations about people types based on demographic variables, it seems like there isn’t more customer segmentation happening based on more behavioral data (or at least proprietary segmentation like that is being kept under wraps). Market segmentation is still almost entirely demand driven, that is marketers come to Acxiom looking for specific parameters to define their customer segmentations, and hasn’t yet evolved to take advantage of the promise of big data to drive segmentation from correlative discoveries in the data.

For example, Howe’s described Porsche looking for the set of consumers who are likely to purchase a luxury vehicle in the next two weeks. The marketers, in this case Porsche, come to Acxiom with the set of parameters and models that will give them that set of customers to market to. Some of Acxiom’s customers are more sophisticated than others as to what parameters they are interested in (i.e. they have data scientists on their teams). So these segmentation parameters are generated by the demands of the marketer. They are essentially hypothesis driven, to match the product to the desired consumer behavior and interest data. The marketer says, “I’m looking for these people, Acxiom, show me where they are.” Acxiom will run those parameters, get rid of the twelve-year-olds who are ogling cars on the Porsche website, and deliver Porsche the men who are looking to buy their next midlife crisis fix so that Porsche can better target advertisement to those ready and willing customers.

We haven’t gotten anywhere closer to letting the data tell us about what kinds of segmentations might be interesting to market to, or even more advanced, to let the data define new market opportunities. This would be the supply-defined segmentation model of the data broker. And it seems like an underdeveloped opportunity for brokers to take on a role in defining markets, with a supply-driven market segmentation derived from the correlations. But it’s also reassuring that the big data promises of correlative discovery hasn’t yet resulted in the creation of new markets. The marketers, for the most part, are still defining the segmentation.

But it’s only a matter of time before defining markets with correlative methods becomes the value-adding, differentiating business of data brokers. Right now, segmentation for marketing purposes is only as useful as the market you think you are targeting. But given what we’ve seen in dragnet surveillance techniques for flagging behavioral patterns, I imagine we will begin to see the industry shift to include both demand (of marketers) segmentation, and supply (of the data brokers) proprietary insights gleaned from amassing and analyzing these huge datasets of both demographic and behavioral details. Right now there’s not enough finesse with correlations to handle false positives and false negatives to differentiate signal from noise. But I expect that will change with time.

I’ve been thinking a lot lately about who gets to define these segmentations and categorizations after reading Ian Hacking on human kinds, especially as the big data promise moves definitions away from humans with power to the largest databases with the best algorithms. When that correlative categorization paradigm shift happens, I wonder about the looping effects (Hacking’s term) of correlative categorizations of consumers. What does it mean for an algorithm to define a market segmentation, compared to a marketer to hypothesize their targeted demographic? And more importantly, what are the looping effects of these marketing segmentations if we as consumers don’t get to explicitly engage with and respond to them? Regardless of how they are defined, these categorizations are all influence with no accountability right now.

Howe briefly touched on the problem that once exposed and editable, consumers might falsify or obscure demographic data in aboutthedata, thus diminishing the value of the data. But Howe countered that if you are a fifty year old man, but you feel younger and more like 39 and you want the advertising you see to reflect that, then there’s not a lot of harm in making that change in the data (of course assuming that advertising is the only intended use of that demographic data). Howe cited that all marketing is aspirational, so no harm in giving consumers the ability to declare their aspirations. Market segmentations have been likened to opinions that marketers have about who they think their customers are. But isn’t data-based decision making supposed to remove the vicissitudes and messiness of opinions? Lying about your age shouldn’t be the only recourse to correct when personal preferences don’t match up with statistical norms in a population. That presumes consumers  know enough to understand the effects of any given data point on their desired advertising outcomes. Right now, we don’t have the ability to understand causality in the uses of data because it’s all opaque and hidden. At a large scale, marketers haven’t been interested in this disconnect because statistically significant patterns have been good enough. But if behavioral targeting is to reach its full potential, as I think Acxiom is invested in, it has to account for how individuals respond to, relate to, and rectify these demographically and behaviorally defined categorizations. 

Howe also talked about the need to trust common sense means of defining inappropriate uses of data (like for insurance, healthcare underwriting purposes) as opposed to regulative measures. To me, it is still very hard to have a productive discussion as a society about what “common sense” is, and should be, we don’t have the means to understand how are data is and is not being used right now, let alone how it should be used. Aboutthedata is a step towards declaring what data exists about us, but we still need better means of understanding how it is used. One step towards that for Acxiom would be to expose at least basic market segmentations that result from their demographic details (but the way Acxiom avoids this now is by saying they are both proprietary within Acxiom, and in large part defined by their marketer customers). I maintain that we can’t really have a common sense conversation about appropriate uses of data (i.e. norms) until we can actually start to trace the data and its uses in everyday practice.

All this leads me back to something I’ve been ruminating on for a while, which is the idea that right now we are in an uncanny valley of targeted advertising. Everything feels a little creepy, we can begin to infer how some of this is working, but it turns out what marketers are doing is still remarkably coarse and not nearly as granular and personally tailored as we think or expect it might be. The creepiest ads we see on Facebook are still based on very coarse demographic categorizations or are retargeted from cookies. Still, when we come across a retargeted ad (if we even know for sure that’s what it is) showing us a shirt that we were browsing for on J.Crew, which we already bought in the store with the company’s loyalty credit card, we feel offended and annoyed because we’ve already taken steps to expose our preferences and our loyalty to the brand, and yet they still don’t understand us in a way that feels right. With little guessing, I can assume that’s because there are silos between browser data, retargeted advertising buys, store records and credit card data. But it feels like we’re already at a point where that shouldn’t be the case anymore. I think we expect the personalization to be more advanced than it actually is. The uncanny valley of targeted advertising lies in the fact that we have some idea just how detailed these things could be, yet we have no means of understanding or confirming what’s going on behind the scenes.

Howe was talking about the potential for developing more of a direct relationship to consumers by building the infrastructure, the pipes, and the connections between all these disparate data sources, such that consumers might be able to say more about what they want, with the expectation of some value exchange. I remain doubtful that consumers are willing to spend time and energy managing those relationships except where they already have a vested interest in maintaining profile information. That’s Facebook’s and Google’s promise (and valuation). And that’s also the catch 22, that because we’ve updated our aboutthedata profile, or use the rewards credit card, that we think we should be getting something back for agreeing to participate in a closer relationship and exposing ourselves, and we aren’t yet.

*This pdf for Personicx seems to be outdated, because the URL reroutes to Acxiom’s new “Audience Operating System” product, but I think this document still gives a sense of the methods behind proprietary segmentation work Acxiom is offering on the supply side of the equation.