What's the use in continuing to run a marketing campaign that doesn't work? It turns out that there might be gold hidden in that dud of a campaign -- you just haven't found it yet. The treasure usually stays buried.
Traditionally, marketers don't have the time or the money to even think about retesting unsuccessful campaigns. In a typical campaign process, marketers decide what to test -- and then determine which is better, A or B? Whether it's a landing page, an email for an outbound marketing campaign, an SMS offer, etc. marketers run tests to measure the impact of certain elements. They wait until statistical confidence is reached. Then they measure the response rate.
If A performs better, it is deemed champion. If a marketing treatment falls below the expected line of performance, it is turned off. In other words, if treatment B doesn't have as much success as A, B is thrown out. Then, 100 percent of the focus shifts to what is working. Marketers launch the "winners" more broadly and hope that the impact is significant enough to move the needle on key business metrics. When campaigns generate negative lift, marketers stop running them, and don't look back.
For example, let's look at a bad campaign. Say a marketer tests two campaigns. One results in a performance lift of +4 percent. The other measures -4 percent. Usually the -4 percent campaign would be dropped immediately.
How machine learning can help
In an ideal scenario, the marketing team would do the analysis to understand which experiences worked for which customers in what context, and why, and then make appropriate decisions based on the insights. But herein lies the challenge. There are limitations to the analysis and testing that marketers can perform manually, preventing the ability to experiment with the full range of potential targeting opportunities that can maximize performance. For marketers, not only is there too much to dredge but in most cases the right equipment doesn't exist to reliably find gold.
This is where machine learning can help. Machine learning can automate functions such as discovery, testing and optimization. For the marketer, machine learning means the ability to learn and the ability to create, based on the learning, the optimal experience for any given customer, and any number of customers, in any given context.
For example, machine learning can dig deeper to discover that within the -4 percent campaign, there are hidden gems. It can discover that there is a subsegment for which lift is +6 percent and another for which it is -10 percent. And that for the segment for which the uplift is +6 percent, there are subsegments for which the uplift ranges between -1 percent and +7 percent. This can go on and on and on.
With recursive experimentation, machine learning can quickly analyze all underlying conditions that determined performance. Maybe the lift - positive or negative - is dependent on whether customers already subscribed to another product, had consumed specific services within the past 24 hours, had recently interacted with friends about the product, downloaded an app or not, were of a certain age, etc.
Modern machine-based technologies offer marketers the ability to test hundreds or thousands of experiences simultaneously, automatically learn what does and doesn't drive lift, and then based on the insights seamlessly and automatically iterate and optimize campaigns. Thousands of attributes and marketing treatments can be explored and measured and the impact that each one or any combination thereof has on performance can automatically be exploited.
It's like a precise, laser focused mining expedition in overdrive. New experiments are continuously executed across behaviors, offers, messages, contexts, etc. so that the machine can learn which user, marketing and contextual elements are causal for positive and negative lift.
Applying this capability to the campaign scenario referenced above, machine learning determines the likely performance result based on any customer attribute or targeting condition, even those not explicitly targeted in the campaign. The assignment of customers to certain offers is then appropriately changed to exploit at all levels, among all subsegments, what is working. In this example, the overall performance increases from +4 percent to +10 percent.
No need to gamble or guess
Often marketers are forced to choose between exploitation vs. exploration.To meet their objectives they feel they have to execute the customer experiences known to maximize performance. However, if they had the time, most would also want to experiment to learn what outcomes may result from other new experiences. It's the classic gambler's dilemma. Do you dedicate your time and money to the one machine in the casino with the highest expected payoff or do you explore the option of increasing your potential winnings by playing new machines you have less experience with? I'll spare you the statistics lesson but we all know that a jackpot only comes so often. On the other hand, add up the smaller wins across hundreds of slots - for which you've determined the patterns for expected payoff - and you're guaranteed to walk away with more coins in your pocket.
Likewise, marketers need to assume this so-called multi-armed bandit approach. They need a system that maximizes performance by executing experiences with the best currently known impact on lift while exploring new experiences that haven't been sufficiently trialed yet to find untapped revenue opportunities. In this scenario, tested experiences that do not perform well at first sight - the dirt, sand and silt - are not discarded. Instead, experimentation on a selective basis continues to hunt for the discovery of gold - those subsegments of customers, contexts, etc. where there may be pockets of lift.
With the ability to dig deep and find nuggets of gold that provide lift for customer segments large and small, marketers can drive considerable improvements in campaign performance. Manual marketing efforts can't scale to explore the insights and combinations of insights that guide continuous marketing optimization, but machine learning applications can.
This is an exciting time for marketers. It's a gold rush, and what's especially promising is that there's enough gold for everyone if they use the right tools.
Dr. Olly Downs is the Chief Scientist behind Amplero, a self-learning personalization platform built by Globys.