Personalised pricing: myth or reality?

Personalisation vs Segmentation

Personalisation is a powerful marketing concept. It appeals to everyone, as it makes us, as consumers, feel more special and unique. Unsurprisingly, service and product sellers use this term broadly to add more perceived value to their offerings. We receive “personalised recommendations” while browsing a marketplace. We get “personalised suggestions” on people to follow while doom-scrolling through X, Facebook, or LinkedIn. We are excited to find a “personalised discount offer” from our favourite store in our mailbox.

According to the survey, that Aimondo team ran in collaboration with the market research agency People Fish, 94% of consumers would consider a seller offering personalised pricing, as the first one in a row while thinking about a purchase.

The problem is that in reality 99% of the offers and suggestions we get are not genuinely personalised. They are actually offers and suggestions that are presented to everyone else who is similar to us in some aspects or behavioural patterns. We, as consumers, are receiving these offers as part of a certain customer segment.

K-Means clustering  

Customer segmentation is not rocket science. Even if you’re not a major seller, you likely do it in one way or another. You might be grouping customers based on their spending history (as part of your loyalty program when they are assigned to a certain tier entitled to a certain level of discount).

Or, you might be grouping customers based on their location and distance from your warehouse to figure out which ones are more lucrative to you, as covering the last mile doesn’t create such a burden on your logistics costs.

A more advanced way of doing this segmentation is to use data analysis technique called K-means clustering.

In general, there are two algorithms that are used to used to segment data into distinct groups based on certain features or characteristics: K-means and K-median. These algorithms are the basis of all “personalisation” that is available today.

K-means Clustering

Is mostly used to segment continuous data

K-median Clustering

Is mostly used to identify outliers.

K-means Clustering is mostly used to segment continuous data. K-median Clustering is mostly used to identify outliers.

• You receive a recommendation on a certain product on Amazon – it’s K-means clustering.
• Your bank transaction to China got banned — it’s K-median clustering at work.

How companies use K-means clustering

As mentioned above, K-means clustering is an advanced technique to segment your customer base.

Another great things that you can do with this algorithm is to analyse Customer Lifetime Value, predict future purchasing behaviour and identify high-value customers.

How to perform K-means clustering (DIY way)

These are the steps you should take to perform K-means clustering.

• Data Collection: Gather relevant customer data.
• Data Preprocessing: Clean and preprocess the data.
• Determining the Number of Clusters (K)
• Interpreting Clusters: Analyse the clusters to understand the characteristics of each segment.

Let’s dive a bit deeper into every step.

Data collection

Examples of data Required for K-means Clustering

• Customer Demographics: Age, gender, location, etc.
• Purchasing Behaviour: Purchase history, frequency, average spend, etc.
• Online Engagement: Click-through rates, page views, time spent on site, etc.
• Customer Feedback and Reviews
• Loyalty Program Data: If applicable.

Data pre-processing

Data pre-processing means that you have to bring it all to a unified format suitable for analysis.

Choosing the Number of Clusters (K)

This is actually the most tricky part. The accuracy of all your analyses and predictions is based on how accurately you choose K.

Imagine this: you’re a salesperson who has just entered a hall where a high school prom is taking place. You observe different groups of people, and you know for sure that there is a certain number of clusters with which you deal. But how many? 2 (teachers + students)? This distinction is helpful for a bartender deciding whom to offer liquor drinks to, but you’re selling stockings and sports shoes. Boys/girls? Great, but you have some trainers suitable for girls who play soccer as well. Maybe you’ll be more successful if you choose 4 as a number for K? Fashionista boys, fashionista girls, sport-loving girls, sport-loving boys? Do you treat teachers as part of the group?... You’ve got the idea.

So, how do serious data analysts choose K? There are several methods that can be helpful.

1. The Elbow Method involves plotting the within-cluster sum of squares (WCSS) against different values of K, and the 'elbow point,' where the rate of decrease sharply changes, provides the optimal K.
2. If this method is too complex try to use one of the distance measuring indexes: Davies–Bouldin Index‍
3. Or Dunn Index

The idea here is to measure distances between “cluster members” and group the ones that are closer to one neighbour than to another. If you still struggle here’s a hands-on article on K-clustering methods that covers all existing approaches

Interpreting Clusters

Final step where your look at your clusters and try to understand characteristics of each.

• What similar behaviour do people show in this clusters?
• What products do they buy most? When?
• What are the triggers for purchase for them?
• How often do they buy these products?
• When are they likely to make a purchase next time?

How to simplify customer segmentation?

Simplifying the process is absolutely achievable! Predictive analytics and BI tools for retail can execute this task effortlessly. At Aimondo, we offer this functionality as a part of our Predictive analytics module. If you're a user simply upload your data, and within minutes, you’ll have your clusters and predictions on purchase intent.

Can personalisation be more personal than segmentation?

To revisit our original question: is clustering the pinnacle of personalised marketing? Will our personalisation only be as good as our categorisation of a customer into a specific segment or cluster?

The answer is yes and no.

Some aspects can be more refined. For instance, you can provide 100% personalised discounts on items to a customer with a distinct purchase history of those items. With Aimondo Flex platform, this is also a simple process: you just select a dynamic pricing strategy that assigns specific discounts to specific products for a specific customer.

However, it’s likely the limitations of clustering will still apply when it comes to predicting behaviour and proposing deals on products that people are likely to buy.

Why? Because a crystal ball that predicts the future does not exist. Predictions can only be formulated by analysing the behaviors of other customers who share characteristics with the one whose behavior you are attempting to predict. Therefore, the accuracy of your predictions will always hinge on the extensiveness of your database, the number of variables included, the accuracy in attributing individuals to specific clusters, and the precision in estimating K…

In essence, it relies on the proficiency of your data analytics team and the quality of the technology you employ.

If you're finding this article useful please join our free Pricing Certification course where we cover K-mens clustering, predictive analytics and price perception in more details.

‍