How do you segment your customers by their online behavior?
People often intend to do one thing but do the opposite. I want to eat out at least five times a week but I only do it once.
I say I want Levi’s jeans, but whenever I go into the store, I never leave with anything.
I say I prefer shopping online, but always find myself browsing for products in brick-and-mortar.
This is called an “intention gap” between what consumers say and their actual behavior.
So how do you close the gap?
If you want to understand your customers’ true purchase intentions, the answer lies in leveraging behavioral data for customer segmentation.
In our previous article on customer segmentation examples, we gave examples of demographic, behavioral, and psychographic segmentation. Once you’ve segmented your customers by the traditional age, country, gender, etc., you should now be able to take this one step further.
These behavioral segmentation examples, tools, and techniques will help you hone your segmented campaigns.
Mastering behavioral segmentation is key to closing that gap between intention and behavior. Moreover, it will help you predict your customers’ future behavior patterns.
This article provides insights from the Director of Conversion and Data Optimization at Microbe Formulas and Crobox’s Chief Data Officer to give you the full picture.
Here’s what we’ll cover:
“If you track what your customers click on on-site, from buttons and links to landing pages, you’ll be able to create more personalized messages across all your channels.” - RoAnne de Weerd, Director of Conversion and Data Optimization
Tracking behavioral data for eCommerce isn’t just about those big-wins in terms of high-impact KPIs. In fact, to properly track behavior, you should look at your micro-conversions above anything else.
Not only will these tiny behaviors reveal your webshop’s bottlenecks, but they’ll also help you see what common behavior patterns people share when shopping online.
So how do you track behavior on-site?
Heat maps show the areas of your site that your customers focus their attention on.
They map on-site activity showing what appeals to certain customers (like subscription, discount offers, even what products are popular for who, CTAs, product categories, etc.).
There are a few different types of heat maps, like:
Just search for any of these online and you’ll see a plethora of eCommerce tools that do this for you.
These are sometimes called “gaze plots” and are similar to heat maps, except eye tracking follows the gaze of the viewer and heat maps follow the mouse behavior. As a result,eye-tracking technology shows the time spent looking, and the chronology of the gaze.
In short, eye-tracking reveals the hierarchy of design: what areas of your webshop draws people’s attention first and last. So if you have a “for sale” category that draws in customers from Spain, you know to optimize this category for that segment. You could also create email campaigns that draw in deal-hungry shoppers.
Not only can gaze plots inform optimization that turns browsers into buyers, but they allow you to segment your customers in a more nuanced way.
All your behavioral data points will be stored in a tracking cookie and can be visualized in tools like Google Analytics.
Make sure your cookies are updated to ITP policies, and (if in Europe) the GDPR, as tracking customers across different sites is becoming increasingly difficult. Focus your personalization in-session as best you can, as this will help track customer behavior in real-time.
Tracking your customers’ transactional data is another straightforward way to collect behavioral data. In other words: how and what people are spending on your webshop.
Make spending categories from high to low, and even spending methods (like PayPal) as these are often static behaviors you can easily segment on.
This works using an RFM model: recency, frequency, and monetary.
Source: CleverTap; check them out for more about leveraging RFM models in their customer segmentation techniques.
RFM is a way to understand your shopper’s behavior based on:
But there’s a caveat: looking at one of these variables isn’t enough. If you have a high-spending customer who bought a lot one-off but never came back, then this shopper isn’t your “most” valuable customer.
It’s the frequent visitors, who spend enough, and keep coming back that will actually be your “high value” segment:
Nevertheless, there’s still value in each of these transactional segments because there are different approaches to engaging them.
Ok, now you have some tools at your disposal for gathering behavioral data. How do you begin to use this data for advanced segmentation?
Here are some behavioral segmentation examples in three easy steps to get you started.
Before testing your on-site behavior, you should already have an idea of who your customers are. Segmentation happens before the fact because it tests your personas and then groups their similar characteristics together. Whereas personalization uses these segments to deliver more tailored customer experiences in the moment.
Whereas personas are micro-profiles and often fictitious, customer segmentation is data-driven and encompasses multi-variable profiles of groups of people.
But before you do anything, you should create your initial shopping personas. For this to work:
Let’s imagine that as a millennial-focused apparel brand, your typical customer persona looks like this:
Your overarching behavioral customer segment would look like this:
“We first build out temporary customer profiles, but this becomes more sedentary compared to when the user has real-life data we can leverage.” -RoAnne de Weerd, Director of Conversion and Data Optimization
Ok, now you have your customer personas. The next thing to do is predict their behavior on-site. This means modeling out a “typical user flow” from the start of their customer journey to the end.
“If you have certain users that engage with a specific geo-targeted campaign, you want to feed those users more relevant content. That’s why behavioral segmentation is so important: It allows you to see who engages with what and how.” - RoAnne de Weerd, Director of Conversion and Data Optimization
For Rita, this could be tracking her click from IG ad to webshop and making assumptions about what would appeal to this same persona as they go through a typical flow or journey.
Since Rita is coming from a high-waisted jeans campaign, it can be inferred that her behavior will filter jeans on high-waist.
Since she’s deal-driven, it’s possible she clicks on the SALE page. And maybe she hovers over reviews on the PDP to make a more sound purchase decision.
All these customer journey trails should be modeled in order to get you to your next step, which is...
Once you have built out your models from any pre-existing assumptions you have about your users’ flow, now you can create different hypotheses to test.
Let’s stick with Rita - although it should be noted that your hypotheses would be less on an individual basis and more on a segment-level.
Your first Hypothesis may suggest: “Rita clicks on SALE products”
In order to make things easier for Rita if this hypothesis were true, you could, for example, give her the option to see high-waisted jeans on SALE. Like through targeted Google ads, or an advanced filtering option.
The more your hypotheses are tested and confirmed by your data, the stronger your customer segments will be.
So if a large amount of traffic is indeed coming from the email campaign into the SALE section, you could:
This is just an example of modeling out a typical user flow from a customer persona. You should always test these flows before rolling anything out. It’s about trial and error, but the more you test the better.
“All the data that comes in from your site is actionable if you can recognize larger patterns within it. If you do have larger data patterns, you know these are outcomes from your hypothesis. You can then identify them as common patterns of behavior for segmentation.” - RoAnne de Weerd, Director of Conversion and Data Optimization
Once you get these processes down to a T, they’ll pave the way for more advanced customer segments.
This is the first step towards predictive modeling for advanced segmentation.
“It’s all about knowing your products and understanding how audiences engage with them. Your entire strategy should be driven by data.”- RoAnne de Weerd, Director of Conversion and Data Optimization
In fact, all the tools we’ve already discussed are a good segway to predict the future behavior of your shoppers. Things like the RFM, testing hypotheses, and modeling out typical user flows.
But as tech-heads, we also highly recommend using AI and machine learning (ML) for advanced segmentation.
A predicted behavior hypothesis you plug into an ML model could look something like this:
Then you test this hypothesis to see if your assumption is correct, in order to anticipate future behavior. Retailers are already using predictive analytics to enhance their customer segments for better targeting - like Target! Pun unintended.
The NYT article that revealed Target’s predictive modeling.
Using predictive modeling, Target found that a teenage girl was pregnant by analyzing her on-site behavior. They then sent her timed coupons in the post at different stages of her pregnancy.
As you can imagine, this didn’t go down well with the girl’s father who found out she was pregnant through Target’s...well, targeting. So here you have an extreme case of predictive modeling using data.
Your own customer segmentation techniques should use customer data in a way that will ultimately benefit the shopper. This means leveraging a customer-centric data-driven approach.
One of the tools to do this is the k-means cluster analysis method.
K-means clustering is an unsupervised machine learning technique that identifies subgroups within your datasets.
Applied, k-means automatically makes behavioral connections between your data points (which are your users, the colorful dots above) to identify similar clusters of customers.
This comes into play after you’ve collected your customer’s behavioral data. The next step is to transform this data into numerical representations (i.e., vectors) and apply the k-means algorithm using ML. Now, this may seem scary for non-ML experts (like me, by the way), but k-means is actually one of the more accessible algorithms.
So how does it work?
The k-means pick random points (or “centroids”), and with every iteration, your users - dots on the graph - will come closer to those centroids.
From App Gate, grouping flower measurements together.
Then, after each iteration, the k-means rebalances all users based on their distance to the nearest cluster until an equilibrium is reached. In the end, you have clusters - or segments - of customers with shared attributes.
A caveat of this is that you'll have to enter manually the number of clusters you end up with. But once you do - there you have an automated customer segmentation technique!
But there’s anothe catch. Because this is “unsupervised”, the clusters that are formed are actually un-classified. Meaning that while you have clusters, you don’t actually have an overarching segment - yet.
That’s when your “supervised” learning comes in or, in other words, where it's up to you to classify your clusters in a way that makes sense for your marketing strategy.
For example, let’s say on your y-axis you have “browser type”, and on your x-axis, you have “device”. Here’s what you do:
From here, you need to classify your clusters. For example, let’s say you have classified three groups of customers based on their behavioral data, looking something like this:
In order to see if these segments are driving purchases on-site, you need to test your communication. Start with A/B testing to find out which segments are indeed driving behavior on-site.
“K-means create clusters, and these clusters are dynamic because they will shift with incoming users along with their unique behavioral data. The only way to create stable segments is to test.” - Leonard Wolters, CDO at Crobox
In short, k-means is just the beginning of your customer segmentation. What follows is to assign those clusters a group, test those groups to see if they work as segments, and then drive the customer experience with personalization.
From CCG Retail Marketing
Let’s talk about personalization for a second. If you really want to personalize the customer experience on an individual level, the k-means algorithm also comes in handy.
K-neighbors are the distance between users instead of the distance between users and clusters. This is the start of personalized segmentation.
Instead of assigning users to a cluster, you can also create “k-neighbors”, and calculate the distance between a targeted user and his/her nearest neighbors.
This would look like: if one user was related to its neighbors because they were looking for similar products (instead of looking for the same product).
This kind of individualized segmentation works for product recommendations. But it's also expensive and time-consuming.
“One pitfall of k-means is a self-fulfilling prophecy. For example, if a group of users is clustered based on their preference for Sci-Fi books, it doesn’t make sense to recommend them Sci-Fi books, since this would be an obvious recommendation for them. You should avoid these kinds of self-fulfilling prophecies by experimenting continuously.” - Leonard Wolters, CDO at Crobox
The k-means is the first step to automating your customer segmentation and also predicting whether customers are likely to buy something similar again.
You can also use the data from your cluster analysis to optimize your RFM tables. Which will help you dive into the customer personas within your segments to predict how they’ll behave in the future.
Advanced customer segmentation starts with behavioral data. This is a data-driven approach to avoiding the intention gap by focusing on what your customers do rather than what they say.
What you’ll have in the end are advanced customer segments; these are important for on-site targeting, retargeting campaigns, getting to know your webshop better including its bottlenecks, and overall understanding your customers on a deeper level.