All marketers want to evoke action from their target audience. We try, in our digital marketing and advertising, to deliver relevant content that causes our audiences to act.
A lack of conversion often corresponds to a lack of relevant content. And lack of relevance is a result of failing to understand that audience’s interests and needs.
Here comes the tricky part. The groups we are marketing to are diverse, as is wont to happen with human beings. So understanding those interests and needs in detail sufficient enough to meet shared goals can be very difficult. Add in the attempt to understand goals at an individual level to drive personalization, and the task of understanding each visitor becomes monumental.
To tackle this problem of delivering relevance in general — and then in a more personalized sense — we should start where we have the best chance for success. Then we can incrementally dig into harder and harder problems.
Old and New Approaches to Grow Conversion
Random testing is how marketers traditionally mine a vein of potential conversion rate growth.
Random testing can manifest in univariate or multivariate forms.
Univariate random testing means we’re testing, or changing, just one variable. These tests are often referred to as A/B tests, though we can change the one variable in multiple ways, which makes it an A/B/C/n … test, for however many variants of change we want to make in our one variable.
Multivariate random testing means we’re changing more than one variable in our page, so we might change a banner and a button CTA. In such cases, we need to test every possible combination of banner and button. If we’ve changed each variable in multiple ways, then our combinations, and thus number of test versions, can get very large.
In both cases, we are conducting random testing, meaning that for however many versions of a page we are testing, we distribute them at random across our traffic, resulting in a sample of visitors that should all look alike for each version. In the end, we compare all the versions in terms of their expected performance if shown to all visitors and pick the version that would work best.
This type of testing has nothing to do with the dream of personalization that haunts every marketer today. Quite the opposite — in determining what works best for everyone, this approach is all about generalization.
Despite the hype, “doing personalization” should not be a business goal, it should be understood as simply another method to achieve a goal of better conversion.
And given the long history of success that random testing has seen in delivering lift, personalization is certainly not the only way to create lift in conversions. In fact, personalization is still a quite unproven method for most marketers.
Proving the Power of Personalization
That personalization's power has yet to be proven doesn't mean marketers should ignore it. Marketers should already be trying to understand how, and how well, personalization can work for them. If you still need convincing, look at it this way: if you believe that continually seeking what is better for all offers opportunity, think of the greater opportunity delivering what is better for each creates.
If the landing page has conversion, then I’ve already found something that works for certain people. If I then run a random test to try and improve upon what I was converting before, I take a portion of these people away from what I know worked for them, and introduce them to a new version that may or may not work for them.
Even if my test version sees lift over control, what I can’t easily tell is whether I had to make trade-offs to get those results. I don’t know whether when I took the prior converters into the test experience, I lost some of my old conversion and had to replace the gap with new conversions before I started seeing lift.
In the end, I may have achieved lift, but in doing so, I’d prefer to ensure that I can keep doing what works with the group for whom I know it works, while testing amongst the groups that weren’t converting. Then, when I raise my conversion rate, I know the increase is coming from additions to what was already working, versus substitution for some group that had been converting but now finds the new experience sub-optimal.
This approach would require us to sort traffic between the type that is expected to convert in the current state, and the type that would not (and so is our target for the test). Sorting of this type is not possible with random testing — it is explicitly non-random — but it is possible with machine assisted audience sorting and content targeting.
Feeding the Machine That Will Feed Personalization
When splitting traffic randomly, we have no interest in the characteristics of each visitor.
Machine assisted sorting takes the reverse view, paying close attention to measurable characteristics around each visitor.
What those characteristics might be will depend on each marketer’s sources (and collecting visitor-level data is an entirely different conversation for another day). But ideally, marketers will have some detail about visitors’ technology and prior engagement with content, perhaps a link to the CRM system for existing customers, and may also have insights into demographics and even attitudes through a Data Management Platform.
For machine assisted sorting, we allow a machine learning algorithm to observe our flow of traffic against eventual conversion or non-conversion to some goal. The more data we can provide about each visitor, the more opportunity we have to let the machine find characteristics that predict whether a visitor will reach the goal.
For personalization to ever work for a marketer’s organization, a fundamental requirement is finding the data that allows a machine to predict conversion or non-conversion within the base current state experience. If we can’t use data to predict conversion versus non-conversion against a single experience, then we will have no chance of using data to select the one of many possible experiences that would best drive conversion for each individual.
Once the machine starts identifying patterns in data that it believes predicts outcomes against goals, we can feed it new experiences to let it try and optimize goals outcomes with people who weren’t converting to the current state.
In a recent six month program our agency ran with a client (from April through October), we introduced three targeting interventions of multiple variations each.
- In the first intervention, the machine discerned a new audience which produced 4.3 percent lift over our baseline
- In the second intervention, lift increased to 12 percent over baseline
- In the third intervention — which offered the most variation to account for the audiences we’d seen emerging through the first two interventions — the collective conversion across the three variations increased 23 percent over baseline
The last result is the product of delivering truly incremental conversions on top of what was already working. Rather than trading what works with one audience for what works with another larger audience, we effectively engage both (all three in this case) audiences with content that is most relevant (and thus, impactful) to them.
Our next step in this program is to begin digging into each of these three distinct high-converting audiences to find sub-segments of motivation within each. The goal here is to further optimize performance within each of three distinct high-intent audiences, as opposed to within one group of ‘all search traffic’ which was the common approach with this (and most) testing prior to machine-learning based targeting.
Do We Really Need 1:1 Personalization?
The path to personalization lies in this process of sub-dividing audiences further and further within prior divisions. At some point in this sub-division of sub-segments, we will reach a segment of one.
Would it be useful to do so? That's not certain.
Personalization of the 1:1 variety may turn out to only apply in certain contexts where recommendations and context-specific content really matters. But for presenting content that motivates conversion to a product or service, we may find that the most productive segments are larger than one, and what we’re seeking in “personalization” may really be “hyper-segmentation.”
One thing that is certain — uniquely addressing segments smaller than “all traffic” provides a clear opportunity.
To develop targetable segments that ensure the highest ROI in our optimization programs, we use consumer insights research around the consumer and their journey to hypothesize what different interests, needs and barriers might exist amongst potential consumers. We then develop what we feel is the best solution, and launch it to validate that, when targeted, it delivers lift to some audience over the prior experience they had.
We're in the Midst of Transformation
With a method for understanding various consumer segments, their barriers as they move through the consumer journey, and a mechanism for sorting audiences into groups that will find each form of engagement more relevant than any other, marketers will be ready to begin their journey toward whatever form of “personalization” will give them the best ROI.
It does not have to be Amazon or Netflix level personalization, but all marketers need to engage their consumers on a more granular basis than the current approach of one large mass. And this will not be achieved by technology alone. It will take changes in marketing culture, organizational structure and process.
Despite the hype cycle around personalization discussed in the first post in this series, the transformation of marketing into a practice of delivering relevance and meaningful engagement along the consumers’ journey to purchase is well underway. And so far, the results of mining for conversion lift through better understanding and targeting of consumers look good.
(Editor's Note: Scot Wheeler will discuss these methods for uncovering consumer interest and shaping consumers’ intent to convert in greater detail at CMSWire's DX Summit this month in Chicago. The conference takes place Nov. 14 through 16 at the Radisson Blu Aqua.)