What we have learned building our Customer Lifetime Value (CLV) models
By Jovan Popovic, Published 11th November 2026
Models that calculate how much a customer is worth to a business have been around for close to four decades, and most in the industry are familiar with the core ideas. One thing I noticed since joining Ellipsis as a Machine Learning Engineer is how differently we talk to CLV compared to classic literature. Whilst the metrics stay the same, the way it's developed here is the gamechanger.
What CLV should measure
CLV is the present value of the future margin a customer will generate1. Most businesses would agree with that definition but, the biggest differences I've seen are in how margin is determined: whether the figure takes out the cost of serving the customer, including what the loyalty program spends on them, or stops at revenue. After that, it comes down to how far ahead you look and whether you discount future value. Each choice makes sense for a different purpose. For loyalty, CLV is most insightful when built customer by customer. Segment views can work too, however, building one CLV average across the base blurs the difference between customers that loyalty decisions act on.
Common shortcuts and where they fall short
I’ve seen three common types of measures masquerading as CLV.
Historical spend. Totals what each customer has spent to date. Useful for reporting, ranking current customers and any analysis that needs numbers anyone can audit, but it shows who was valuable, not who will be.
Revenue-based projection. Projects revenue forward for each customer or segment. Useful for top-line planning and for businesses that don't track cost data by customer, but without costs it overstates what a program adds.
Single average. One value applied across the customer base. Useful for high-level budgeting and quick estimates of what can reasonably be spent to win a customer, but it hides the spread between customers that loyalty decisions act on.
Loyalty decisions need more detail because they involve choosing between customers, spending money on rewards, and measuring whether behaviour changed. The useful comparison is the profit left after that spend, and that places different demands on CLV.
What a powerful loyalty CLV model does
Uses margin. We deduct product returns, service costs and loyalty program costs such as points, rewards and administration. A loyalty program is an investment in customers, so the number that matters is the profit left after that spend. Without program costs, the value attributed to a program is overstated.
Looks forward. Loyalty investment is aimed at future behaviour, so we predict future value from purchase behaviour, at the level the data supports.
States the horizon and discount rate. A defined end point and discounting let customers and investments be compared on the same basis: today’s dollars.
Keeps the spread in value. Customer value is uneven, with a small share of customers typically carrying a large share of the total2. Tier thresholds and reward spend choose between customers or segments, so the spread matters more than the average.
Tests against actual behaviour. We compare predicted value with what customers went on to deliver. Where the two diverge, we adjust the model, improving it iteratively.
Why CLV isn't One-Size-Fits-All
Good CLV modelling starts with one question: how does this business relate to its customers? The answer tells you which model to use. Below are three commonly used model families and the industry to use them in:
Retail customers rarely tell you they've left; they just stop coming back. Buy ’Til You Die models suit this (also called ‘Always a Share’ models). They use how recently and how often each person has shopped to estimate the chance they're still an active customer.
Fuel (and most consumer goods) is closer to ‘always a share’. Customers fill up wherever is convenient, split their spending across brands (Frosted Flakes one week, NutriGrain the next) and drift in and out without ever really leaving. The value there is in how much of each customer's spend you win over time, not whether you've lost them.
Contractual businesses like gyms and phone companies sit at the other end of the relationship dimension, in a ‘lost for good’ relationship. Customers leave in a visible way by cancelling, and once they're gone, they're treated as gone, so value comes down to how long members stay and how likely they are to renew.
Fitting the right model is key in matching the correct assumptions about your customer base with the business model you operate.
The data also must support the model. The most useful CLV is calculated for each customer because it drives targeting, personalisation, and spend decisions. We push for the most detailed view the data can reliably support. But in businesses where customers buy rarely, such as car sales booked every year or two, there isn't enough repeat behaviour for a customer-level model to learn from. Sparse data can conclude, with confidence, that nobody ever leaves. Rather than deliver a number we don't trust, we group customers into behavioural segments and value each segment using the approach proposed by Prof. Gupta3, which is based on its retention rate and the margin it generates. Choosing the right model for the business is key to making it a metric that can produce actionable insights.
The value is in how the number is used
On its own, a single CLV model is simpler to build than it used to be. The tools are mature and the techniques are well documented. The effort that matters is the profitable application of the output.
For us, CLV sits within Ellipsis's Return on Loyalty® (ROL) platform. Across our client work, the most engaged customers tend to join a program first, so comparing members with non-members credits the program with value those customers would have created anyway. ROL corrects for that, and CLV adds the forward view: what the margin a program adds is worth over a customer’s lifetime, not just this year. A program earns a return when that added margin outweighs its cost to run. Getting CLV right is not a spreadsheet exercise. It takes a specialist team of data scientists and ML engineers, plus modern, secure data infrastructure that can bring together years of transaction, engagement and program data and model it properly. That is why Ellipsis has invested in the Ellipsis Data Cloud™, a secure, scalable, AI-enabled platform, ISO 27001 certified, that runs our proprietary Return on Loyalty® models, including CLV Pulse and Loyalty Attribution. Plenty of people talk about CLV. Far fewer have the people, the platform and the methods to measure it in a way a CFO will trust.
Our clients use CLV to inform four decisions.
Tier thresholds. Where the lines sit and who qualifies.
Reward spend. Customers who have already spent close to their predicted value have little room to grow, so extra rewards mostly pay for spending that would have happened anyway. Customers with a large gap between what they've spent and what they're predicted to spend are where rewards can change behaviour.
Retention targeting. Which at-risk customers are worth the cost of keeping (those with the highest future value).
Acquisition payback. How much should you invest to win a new customer. As an illustration, if predicted CLV for a segment is $800 and the target return is 3 to 1, the most worth paying to attract a new customer is about $267.
CLV earns its place in decisions
CLV has an optics issue, where, when used arbitrarily, it can water down the value that these metrics produce. Commonly seen as a forward-looking figure that helps companies estimate how much their existing customer base may spend in years to come. Rather it should be viewed as decision making helper for marketing and financial teams, by tracking the estimated spend of a customer through their interactions with the client. CLV values per customer will shift as data evolves, but this shift helps raise red flags early, change priorities in who to retain and where to spend the budget more effectively. Whatever version a business uses, its value shows in the decisions it informs and the profit those decisions return. In loyalty, those are how tiers are set, where reward spend goes, who is targeted for retention and what is paid to acquire customers. The model is a small share of the work. Most of it goes into the definition, the data and the decisions that follow. Our CLV is built for those decisions.
We are Ellipsis, The Loyalty Experts®. We help you find, understand, measure, manage and grow customer value. We’re here to help, let's talk.
References
Gupta, S., Hanssens, D., Hardie, B., Kahn, W., Kumar, V., Lin, N., Ravishanker, N. and Sriram, S. (2006). “Modeling Customer Lifetime Value”. Journal of Service Research, 9(2), 139-155.
Fader, P. (2020). Customer Centricity: Focus on the Right Customers for Strategic Advantage (2nd ed.). Wharton School Press.
Gupta, S. & Lehmann, D.R. (2003), “Customers as Assets”, Journal of Interactive Marketing, 17(1), 9–24; Bendle, N. & Wang, S. (2023), The Customer Asset.
Related Articles
