Imprecise methods for influencing customers and measuring impact have hobbled marketing in an increasingly data-driven era. A breakthrough discovery may be about to change that.
In May 2017, a computational social scientist from The Psychometrics Centre at the University of Cambridge stood before an audience at the Linux Foundation’s Apache Big Data conference and revealed how close we’ve come to the ultimate goal of marketing: an easily scalable, highly accurate way to predict customer preferences using minimal data.
When she was still a PhD candidate, Sandra Matz created a Facebook ad campaign targeting people based on nothing more than how extroverted their Facebook Likes indicated they were. People with Likes associated with extroverts saw ads for a party game played in a group. People with more introverted Likes saw ads for a quiet game meant to be played solo.
The campaign required only simple algorithms and no advanced analytics. Yet over seven days of testing, the targeted ads generated up to 15 times higher click-through and conversion rates—and significantly more purchases and revenue for the game company.
“We developed this approach to show that you can achieve highly accurate behavioral and psychological targeting with a minimal amount of data and relatively simple machine learning tools,” says Matz, who is now an assistant professor of management at Columbia University’s business school.
As effective as this experiment was, Matz suggests that it’s still rudimentary compared to what could be done with more and richer data from more sources. And it’s downright primitive given the possibilities of applying more sophisticated Big Data analytics.
These possibilities have created a watershed moment for marketing and its role in the business.
Spiraling Down the Marketing Funnel
Tension has always simmered over marketing’s contribution to business success. The business knows it can’t sell products or services if it doesn’t make customers aware of them, but the impact of marketing strategy on sales and revenue is hard to quantify and reliably replicate—which, in the age of the data-driven enterprise, often leaves some business leaders not just undervaluing marketing but actively mistrusting it. No wonder human resources consultancy Russell Reynolds reports that the 2016 turnover rate among CMOs was the highest it has seen since it began tracking the statistic in 2012.
Most companies still determine customers’ readiness to buy by using a primitive model known as the marketing funnel, which sorts customers into increasingly smaller groups as they progress from first becoming aware of a company to buying, using, and finally advocating for the company’s products. Different versions have different definitions and numbers of stages, and some approaches see the model as a circle, but they all have one thing in common: their ability to sort customers into various stages is limited by the amount of knowledge the company has about each customer.
As a result, the marketing funnel ends up leaking. Some customers back away because they feel harassed by campaigns that don’t apply to their needs, while some of those who are interested fall through the cracks from a lack of attention. Many data-hungry business leaders think of the marketing funnel as no more than a variation of “throw something against the wall and see if it sticks,” and with the proliferation of digital channels and diffusion of customer attention, they have less patience than ever with that approach.
The silver lining is that a more precise, quantifiable way to build customer relationships is emerging. Done properly, it promises to defuse the tension between marketing and the rest of the business, too.
The Defining Moment
The Cambridge University experiment is one more step toward the long-held marketing dream of the “segment of one.” This concept of marketing messages that are highly granular, even individually tailored, has been around since the late 1980s. Over the last 15 to 20 years, as customer behavior has become digitalized as never before, marketers have been optimistic that they could capture this data and use it to tailor their messaging with laser-like precision.
Yet what’s achievable in theory has been impossible in practice. We’re still struggling to find the right tools to move beyond the basics of demographic targeting. The rise of the internet, smartphones, and social media has generated more types of information about customer behavior in larger amounts than ever before. But using digitally expressed sentiment about everything from toys to turbines as the basis for accurately disseminating highly individualized marketing messages is still time consuming and cost prohibitive.
The algorithm needed just 65 liked Facebook Pages to know someone’s personality better than their friends do.
However, experiments like Matz’s are bringing us closer to creating highly personalized customer experiences—perhaps not always at the individual level but certainly at a level of granularity that will let us unequivocally determine how to best target and measure marketing programs.
Liking Lady Gaga
Between 2007 and 2012, Psychometrics Centre researchers gathered seven million responses to a simple questionnaire for Facebook users. The carefully designed questions measured respondents’ levels of extroversion, agreeableness, openness, conscientiousness, and neuroticism, a constellation of basic personality traits known as the Big Five.
With the respondents’ permission, the researchers used simple machine learning tools to correlate each person’s responses with the official Facebook Pages that the person had liked, such as Pages for books, movies, bands, hobbies, organizations, and foods. They soon saw that certain personality traits and certain Likes went hand in hand.
For example, most people who liked Lady Gaga’s Page tested as extroverts, which made liking the Lady Gaga Page a relevant data point indicating that someone was probably an extrovert. By 2016, Matz was able to create a lively Facebook ad to be shown only to people who had liked a significant number of official Pages that seemed to be linked to extroversion. A more serene ad was shown only to those whose Likes suggested that they were introverts.
Despite the large size of the Psychometric Centre’s data set, what’s most remarkable about its work is how few data points within that data set were necessary to build a reliable profile that could model useful predictions. Matz told EnterpriseTech that the algorithm the Centre developed needs, on average, just 65 liked Pages to understand someone’s Big Five personality traits better than their friends do, 120 to understand them better than their family members, and 250 to understand them better than a partner or spouse. This may be the first sign that the era of true behavioral marketing is upon us.
Of course, most marketers want to know far more about customers than how outgoing or reserved they are. Scraping Facebook Likes isn’t enough to give them the holistic customer understanding they crave—not when they have an entire universe of other data to consider. The race is on to identify from the vast spectrum of available customer data not only which specific online behaviors have a predictive element such as extroversion or introversion but also which ones will drive the most potent response to specific product or service messaging.
Complicated? Yes—but we are within reach of the algorithms we need to connect the dots for greater customer insight. By reaching out over new channels with more accurate behavior-based messaging, companies could transform the entire customer journey.
A Customized Journey for Each Customer
Attribution, the ability to know the source of a sales lead, is key to behavioral targeting. The more details a business knows about what its customers have already done, the more accurately it can predict what they will do next.
In the past, developing a customer profile relied on last-touch attribution analysis, that is, evaluating the impact of the last interaction a prospective customer had with a brand before becoming a lead. The problem was that companies could rarely be certain what that last touch was, given how much activity still takes place offline and isn’t captured or quantified.
Companies also couldn’t be certain how, or even if, a last touch—be it downloading a white paper, visiting a store, or getting a word-of-mouth recommendation—accelerated the customer through the marketing funnel. They could only predict revenue by looking at how many people were deemed to be at a specific stage and extrapolating from past data what percentage of them were likely to move ahead.
Today, we’re capturing so much more information about people’s activities that we have a far more accurate idea of both what the last touch was and how influential it was. Behavioral targeting makes any content a customer interacts with valuable in analyzing the customer’s journey. A company can use hard data about those interactions to see where each individual prospect is in the customer journey and predict how likely each one is to continue moving forward.
The company can then generate a tailored offer or other event to nudge individuals along based on what has been successful with other customers who buy the same things and behave in the same ways. For example, a large grocer may send out two million individualized offers each week based on loyalty card activity. This may not strictly create a segment of one, but it creates many small segments of customers with similar behaviors based on what the grocer knows to be effective.
As Cambridge University’s experiment in creating an algorithm to identify and target introverts and extroverts proves, more precise messaging is more effective. By using more complex machine learning algorithms to further filter and refine successful messages to target smaller groups, companies could boost their conversion rates to as high as 50%—an exponential increase beyond today’s average rates.
By using machine learning to speed up the testing of different campaigns and to continuously compare results, companies could rapidly create a dataset about every potential customer’s responses and then benchmark it against others’ responses. This would let them determine individual prospects’ likely responses based on concrete actions rather than assumptions.
For super-luxury brands with a limited number of customers and the ability to capture a vast amount of information about each one, this could lead to true segment-of-one marketing. For other brands, the challenge is not just to figure out who the customer is and what messages to send but also how to scale that personalization to segments of tens of thousands (or hundreds of thousands) of customers at a time. To do that both effectively and quickly, companies will need to leverage machine learning, the Internet of Things, and other advanced technologies that enable accurate predictive models. Companies can then benchmark their projected hit rates against their actual results and refine their algorithms for even greater agility and responsiveness.
The Next Steps of Predictive Marketing
Effective behavioral targeting requires companies to identify all the relevant data points, including external data points that indicate which information is valuable. This calls for data scientists who can spot and remove the irrelevant data points that are at the far ends of the curve and distill what remains into meaningful algorithms. It also requires machine learning tools capable of high-volume, high-speed listening, assessing, learning, and making recommendations to improve the algorithm over time.
Once you’ve created a baseline of primary criteria, you can determine the important criteria by which to segment your customer base. To use an oversimplified example, imagine that you own a coffee shop and you want to increase sales of high-margin bakery items. You need to look not at the customers who always get a muffin with their coffee or at those who never do but at those who buy a muffin sometimes, so that you can start to identify the triggers that make them choose to indulge.
Marketing could become about informing customers about their options at any given moment, in real time.
To scale this process, look at both user-based and item-based affinities. User-based affinities link customers who have similar interests and shopping patterns. Item-based affinities link customers based on what they buy, individually or in groups of items. Using machine learning to pair and cross-reference these two factors will enable you to create messages that are personalized enough to seem individualized, even though they’re actually targeting small, multi-person segments.
Retailers of all types collect data about individuals, down to location, date, time, and SKU of the sale. They may experiment with behavioral targeting by making in-the-moment offers based on what they already know about their customers. For example, they may use a mobile app with geofencing to be alerted when a customer using the app is in the store. The alert triggers back-end systems to look up the customer’s purchase history, generate a relevant offer, and deliver that offer to the customer’s smartphone while the customer is still in the store.
The Line Between Marketing and Manipulation
Just the idea of receiving marketing messages influenced by their behavior will disturb some customers. When marketing is designed, as behavioral targeting is, to maximize engagement, the value of the content depends less on whether it’s useful to the audience or even true and more on whether it gets the target audience to engage and reveal another piece of the behavioral puzzle. As a result, companies considering behavioral marketing must consider a question as old as marketing itself: where is the line between advertising and propaganda?
Creating personal profiles of customers based on their actions and personalities will become inexpensive and easy, for better or worse. Better will lead to more relevant and compelling offers based on predictive models of what customers would like to buy next. Worse will create (or at least look like) scalable, granular manipulation.
If companies hope to apply this level of targeted marketing without coming across as intrusive or invasive, they will need to be completely transparent about what they’re doing and how—and with whom they’re sharing the information. Most shoppers say they’re willing to give up data about themselves if it leads to a better shopping experience and more relevant recommendations.
Numerous studies show that customers are comfortable sharing their buying patterns and preferences as long as it doesn’t compromise their personally identifiable information. Nonetheless, they may decide otherwise if they believe that by welcoming you into their lives, they’re throwing open the doors to strangers as well.
As data mining for behavioral targeting becomes more common, companies will have to offer customers the opportunity to opt in and out at varying levels of detail. They will also need to identify and flag the significant minority of customers who prefer not to be profiled in such depth (or at all). Machine learning will be invaluable in responding to complaints on social media, tracking the relevant details of offers that were ignored or got negative reactions, and otherwise ensuring that companies don’t misuse customer data or misunderstand consumer wants and needs.
“The entire paradigm of targeting and campaign implies a vendor doing something to customers,” says Mark Bonchek, founder and “chief epiphany officer” at Shift Thinking, a Boston-based consulting firm that helps companies pursue digital transformation. “It implies getting people to do what you want them to do rather than helping them do what they want to do,” he says. “Be clear on the mental model behind your behavioral targeting. Is it more like a friend figuring out the right gift for a friend or a salesperson trying to close a deal with a prospect? People don’t want to be targets.”
Instead, Bonchek suggests, think of behavioral targeting as a way to build a reciprocal relationship that lets you enhance the customer experience at multiple touch points, not all of them actual transactions. Utility companies send customers information about their own and their neighbors’ energy use so they can benchmark themselves. The utilities often follow up with suggestions about how to save both power and money. Meanwhile, a credit card issuer could help customers understand their purchasing patterns and discover new stores or service providers.
“Loyalty is an emotion first and behavior second,” Bonchek says. “It’s the difference between pushing customers through a funnel and helping them achieve a shared purpose.”
The Art of Scientific Marketing
In mid-20th century New York City, a small local chain of markets developed a national reputation for customer service. It let favored customers call in orders and pay for them at pickup. Managers kept lists—handwritten lists, no less—of their best customers’ preferred products and called those customers with special offers. People were happy to pay slightly higher prices overall in exchange for exclusive bargains and highly customized service.
Although it leverages new technologies like machine learning and Big Data, behavioral targeting will in many ways bring us full circle to that hands-on era in which companies created relevant offers that made customers feel valued and understood. Matz believes it would be a competitive advantage for companies to let customers interact with their profiles and even correct them to ensure that they only receive offers that meet their needs and preferences.
As more situational data pours in from smartphones and wearables to be analyzed by AI, she adds, behavioral targeting could become something more immersive than mere marketing. “If you know from that data that someone is not just an extrovert with specific preferences but that they’re currently in a good mood, you can start fine-tuning messages for that particular point in time,” she says. “We’ll move beyond static profiles to interactions based on characteristics that fluctuate.”
With enough data to work with, she suggests, behavioral targeting could become less about making offers and more about informing customers about their options at any given moment, in real time.
Denise Champion is vice president of Strategy, Research, and Insights for Global Marketing at SAP.
Jeff Harvey is global COO, SAP Analytics & Insight at SAP.
Lori Mitchell-Keller is global general manager, Consumer Industries at SAP.
Jeff Woods is global COO, SAP Leonardo, Data and Analytics.
Fawn Fitter is a freelance writer specializing in business and technology.