Acquire more, retain more! 

Demystifying customer segmentation

There are stages of maturity when it comes to segmenting your customers. Taking a step by step approach will get you faster to value than jumping straight into advanced AI without a plan.

Stage 1: Using pre-made segments

The first step in getting value from segmentation is one that any marketer can use without the guide of a data analyst.

Start getting acquainted with the attributes already available in the platforms you use.

💡  An attribute is a data point or quality of a customer. It could be a simple demographic like their country or age. Or a calculated metric like how many purchases they've made with you over a specific time period. 

For example, if you're using Google Analytics to analyse web users, you're probably already familiar with being able to segment users by their device type and traffic source

This will provide you a wealth of insight into which metrics are different for certain devices. Is mobile bounce rate higher than desktop? This could indicate the mobile experience is suboptimal, or that when customers are accessing your site on their phone they’re in a different mindset and looking for different content.

Sample breakdown of Platform and Device type in GA4

Stage 2: Rules-based segmentation


The second stage of getting more advanced is rules-based segmentation. This is where you infuse business and domain knowledge to create a set of custom rules in which to segment customers. These rules can be easily scripted using a CASE WHEN statement in SQL (or similar) so helps to have a data analyst.


How you’d usually go about this is to first decide on the goal (how this segmentation will be used), then come up with a list of attributes that relate to this goal.


These attributes are chosen based on any of the following reasons:


To get the most value out of this requires time spent upfront in defining your goals for the segmentation. Are you interested in segmenting the whole customer base to identify the most valuable customers? Or does your strategy mean focusing in on buyers of a particular category, and segmenting them based on more specific behaviours in order to market to them better?


You focus area makes a difference, since different behaviours may be driving sub-segments of your customers.


Case study #1: Acquiring more app users

Say your goal is to acquire more app users by encouraging existing loyalty members to download your new app. Specifically, you want to target members via email


Because your target is existing members, you likely have a wealth of information already captured on their behaviours and preferences.


The below attributes could used to segment members, based on their relevance to identifying those members most likely to download the app:


After building this segmentation, you can use it to improve your marketing targeting, by focusing on the ‘higher quality’ segments (those that already showed higher penetration of app users).


This means sending less emails to achieve the same or greater number of new app users. It can also save costs if you personalise the offers, targeting a higher value reward to those segments who need more of an incentive to get the app, and little to no reward to those segments who simply need to be made aware of the app and are happy to download and try it.

Stage 3: Using modelling algorithms

The third, more advanced stage, is building a segmentation model off the data. This is when you use an algorithm that either: 

The power of modelling is that now you can expand out the attribute set to review 100’s or 1000’s of attributes, and the model does the work of sifting through and finding which are the most useful attributes for segmenting your customers.

It can be tempting to skip the above stages and jump straight to this more advanced stage. But in this stage it's easy to start getting lost in the data. There's a reason why after investing in a Data Science and AI capability, x% of marketers say they're not seeing results.

Think of machine learning like a toddler - it can learn very fast, but it needs more and more data in order to do so. It also needs a little push in the right direction if it's missing the right context or outputting results that either don't make sense to the business, or are interesting but can't be actioned on.

By working through Stages 1 and 2 beforehand, you and your team have naturally been building the context needed to design and train more actionable models.


The model's output

Once a model is built, usually there are only a few hundred of attributes it deems useful.

Any more than this and you risk overfitting the dataset you used to build the model off.

💡  Overfitting is like when you study for a test by memorising the answers to questions from past papers of previous years. Then on test day you come across a completely new type of question and become stumped. 

In a similar way, overfitting in a model means the model can almost perfectly explain the customers and behaviours in your original (historical) dataset, but when you try it apply it to future data (new customers come in or the behaviours of existing customers change), the model won't perform as well.

With enough research, anyone can summarise what happened in history - but that doesn't make us any better at predicting the future!

This is why it's a fine balance to design a high performing model, that maintains enough flexibility to perform just as well on new data. Doing this well is both art and science and where a Data Scientist's expertise is extremely valuable.

Case study #2: Purchasers of a clothing brand

Say your goal is to understand the customers who buy from your clothing brand.


You can create a whole list of attributes you think might be useful to throw into a clustering algorithm like k-means:


The algorithm will try different combinations of attributes to see which lead to the biggest differences between customers e.g. do customers who buy lots of dresses show very different behaviour to those who buy from a range of categories?


With little guidance on your part (apart from the input data, which involves its own work), the model may output a few segments that look something like this:


This will usually prompt some useful and insightful discussion within your teams. 

A segmentation like this can then be used to inspire marketing comms, however the reason this sits at Stage 2, is because it's easy to build a segmentation like this, but then not know how to execute on it. Sometimes a segmentation like this is built more for the discussion it provokes, rather than being the segmentation used to target marketing comms.

Stage 4: What about 1:1 personalisation?

You may be thinking this segmentation stuff all sounds well and good. But what about the 1:1 personalisation everyone's talking about? Isn't marketing to a segment still a blanket-type approach?

This is where the most advanced stage comes in. When you've gained all the efficiencies, optimisations and learnings you can from your journey through Stages 1 to 3 - then you are ready to dive into these advanced techniques.

Stage 4 is really when you need to complement your Data Analyst, by hiring a Data Scientist. The Data Analyst will provide the Data Scientist with the deep domain knowledge and key attributes known to predict behaviour.

The Data Scientist will then feed this into a propensity or uplift model, where each individual customer is assigned a score, relating to their likelihood to perform the desired action given your company's targeted marketing technique.


Case study #3: Acquiring more app users

Let's extend the first case study. We left there with a rules-based segmentation allowing us to narrow our marketing focus on the segments most likely to respond to our 'download the app' message.

 Say your Data Scientist takes the knowledge gained from this segmentation, includes 100s more attributes, and builds a model to score each individual loyalty member on their likelihood to respond to the message.

This time we run the campaign by targeting the top X% of members likely to respond, rather than targeting the top X number of segments.

If the model works well, you should see an increase in app take up if you were to A/B test the campaign and choose half of the audience based on the rules-based segmentation approach, and the other half based on the uplift model approach.

This uplift is due to the 1:1 nature the model is able to deliver. For example, the model will be able to pick up the 'best' customers from even the 'worse' rules-based segment. It's able to do this because it is treating each customer as their own person, rather than as a group of people.

Final takeaway

There are three stages of segmentation maturity and working through each one in order will improve your data literacy, speeding up the time to value.

There's also a bonus fourth stage which is where the Data Scientists come in. But you really have to have your house in order before diving into that.

Have any segmentation journeys of your own? Drop me a messaged on LinkedIn and let's discuss.