The title of this article says “affects” rather than “increases”, and that is deliberate. Personalisation can raise customer lifetime value. It can also raise the appearance of it while doing nothing, and in some circumstances it can reduce the real figure while every visible number improves.
This guide covers the economics: what CLV actually measures, the three drivers personalisation can move, why the most-cited evidence in this field does not establish what people think it establishes, which costs a credible calculation has to include, and how to build a business case that survives scrutiny.
It does not cover campaign design. For the journeys themselves, with triggers, decision rules, risks and evaluation methods, see our guide to personalised engagement tactics that increase customer lifetime value.
What customer lifetime value measures
CLV estimates the present value of the future profit expected from a customer relationship over a defined horizon.
Three words in that sentence do the work. Estimates, because it is a model output rather than an observation. Profit, because a revenue-based figure is a different and much larger number. Defined horizon, because the answer changes with the period you choose.
There is no single universal formula, and treating one as universal is where most CLV work goes wrong. Gupta and colleagues set out the modelling choices in a review that remains the standard reference: models can deduct acquisition cost and the cost of serving the customer, and they vary substantially depending on the setting (1).
The most consequential distinction is contractual versus non-contractual.
In contractual settings, the relationship ends observably. Someone cancels, lets a subscription lapse, or does not renew. You know.
In non-contractual settings, silence is ambiguous. A customer who has not purchased in three months may have defected or may simply be between transactions. Fader, Hardie and Shang describe this as the central identification problem: distinguishing customers who have ended the relationship from those in the middle of a long hiatus (2). Every non-contractual CLV number contains an assumption about which of those is happening, and that assumption is usually invisible in the reporting.
Ascarza and Hardie make a further point worth carrying: usage and churn are related, and modelling them separately can discard predictive information or impose a separation that does not exist in the data (3).
A practical formula, and what it assumes
A general expression:
CLV = Sum of expected future customer contribution, discounted to present value − Acquisition cost
A simplified historical calculation:
Historical customer value = Total customer contribution margin to date ÷ Number of customers
The second is not a forecast until you add assumptions about future behaviour, and labelling it as one is a common sleight of hand.
Whatever form you use, four choices have to be disclosed:
- contractual or non-contractual relationship;
- historical, predicted or incremental CLV;
- whether dropout is observed or inferred;
- which assumptions govern future purchasing, margins and survival.
Change any one and the number changes materially while the label stays the same.
CLV against acquisition cost
The CLV-to-CAC ratio is the standard unit-economics comparison, and it distorts badly when the two sides are computed on different bases.
A fully loaded CAC set against a revenue-based CLV will flatter the business. A short-period CAC set against a mature-customer CLV estimate will flatter it further. Both sides need the same customer population, the same period logic and the same treatment of costs.
There is no universal target ratio. Appropriate economics depend on gross margin, cash flow, payback period, growth stage, capital constraints and how much you trust your own forecast. A high ratio can indicate strong economics, or it can indicate that acquisition spending is too timid, or that costs were never fully counted.
For the formula, required inputs and interpretation warnings on both metrics, see our marketing metrics reference.
The three drivers
In a simplified customer-behaviour model, personalisation can affect CLV through three principal levers: contribution per purchase, purchase frequency and relationship duration. The resulting value still depends on cost to serve, retention expenditure and how future cash flows are discounted, all of which are covered further down.
Contribution per purchase rather than spend per purchase, because a larger basket can produce less profit. Naming which lever a programme is meant to move, before building it, is the difference between a strategy and a collection of campaigns.
Contribution per purchase
Personalisation can raise order value through relevant cross-sell, complementary recommendations, bundles and upgrades.
Bundles in particular are underused relative to how well they work. Three related products offered together, ideally with a discount for taking them as a set, makes the decision easier and the purchase larger. It is unglamorous next to a recommendation engine and I watched brands consistently underinvest in it while spending heavily on more sophisticated things that moved less.
But order value is the driver most likely to produce a number that is true and useless. AOV includes only completed orders, so it rises when lower-value customers stop buying. And incremental product revenue has to be reduced by cost of goods, promotional discounts, returns, support and fulfilment before it means anything.
Shah and colleagues examined five company databases across consumer and business markets and found that between 10% and 35% of customers who cross-bought were unprofitable, and that those customers accounted for 39% to 88% of total customer losses (4). The mechanisms included limited total spending despite buying across categories, revenue reversals, excessive service requests and promotion-driven purchasing.
Cross-sell rate is therefore not a proxy for profitability. It can move in the opposite direction.
Purchase frequency
Personalisation can raise frequency through replenishment timing, next-best-product relevance, and reducing the effort of a repeat purchase.
The evidence that a targeted intervention can causally change behaviour does exist. A randomised field experiment involving 30,297 email recipients found that manipulating name similarity between sender and recipient increased opens, clicks and donation amounts (5). That is a real causal campaign effect. It says nothing about repeat behaviour, retention or lifetime profitability, which is the pattern across most of this literature.
The frequency driver also has a specific failure mode: acceleration without addition. A well-timed reminder can pull a purchase forward without increasing total purchasing across the year. That looks identical to growth in a monthly report and is worth nothing.
Relationship duration
This is the driver with the most attached mythology, and it needs the most careful handling.
The retention statistic almost everyone quotes is misquoted. The claim that a 5% improvement in retention increases profits by 25% to 95% traces to Reichheld and Sasser’s 1990 Harvard Business Review article, which appears to have said 25% to 85% (6). Bain was publishing the higher range by 2000 (7) and Reichheld repeated it in 2001 (8).
More importantly, Pfeifer and Farris showed in peer review what the figure actually refers to. The reported increases apply to the discounted value of customer cash flows, not to company profit (9). The exercise adds roughly five percentage points to the retention rate, so 80% becomes 85%, which is not the same as improving 80% by 5% to reach 84%. And it is a comparative accounting model in which retention changes while margins, retention spending and other inputs are held fixed. In an actual business, better retention usually requires additional service, incentives or discounts, and the customers you retain differ from the ones who would otherwise have left.
Long relationships are not automatically profitable. Reinartz and Kumar analysed three years of daily customer data from a catalogue retailer and concluded that long-life customers are not necessarily profitable customers (10). Tenure and profitability were related imperfectly. Some long-tenure customers remained low value throughout.
The economically accurate proposition is narrower than the popular one: retention creates value only when the discounted incremental contribution from the relationship exceeds the incremental cost of retaining and serving it.
Targeting the highest churn risk is often the wrong move. Ascarza’s field experiments found that customers identified as having the highest risk of churning are not necessarily the best targets (11). A customer’s untreated probability of leaving is a different quantity from their responsiveness to an intervention. Retention spend aimed purely at risk lands disproportionately on people who cannot be persuaded, and on people who were going to stay anyway.
The mechanic that changes the incentive
Across ten years of watching loyalty structures, the businesses that got what they wanted from them were consistently the ones billing customers monthly.
A subscription creates something no points scheme replicates. When someone is already paying you every month, they feel they should buy from your page rather than a competitor’s, because they have already committed. The payment itself does the work that a loyalty programme spends years and a great deal of reward liability trying to manufacture.
That is not a recommendation to bolt a subscription onto every business. Most cannot. It is a lens for evaluating loyalty mechanics: ask what genuine incentive the structure creates for the customer, and be honest when the answer is that it creates one for you and an administrative burden for them.
Why correlation does not prove incremental CLV
Customers who engage with personalisation have higher lifetime value. This is true almost everywhere and it proves nothing, because more engaged customers are more likely both to use personalisation features and to generate future value regardless.
Netflix states the underlying issue directly: most machine learning used in personalisation and search is purely associative, and a simple calculation overstates incremental lifetime value when some people exposed to an intervention would have subscribed anyway (12).
A credible incremental CLV claim needs, at minimum:
- random assignment or a defensible causal identification strategy;
- analysis by assigned treatment, including customers who never engaged;
- contribution margin rather than revenue;
- treatment costs, discounts, service costs and adverse effects included;
- a long enough observation period, or a validated model connecting experimental outcomes to later cash flows;
- uncertainty intervals and sensitivity tests on the retention and discount-rate assumptions.
Comparing people who used recommendations with people who did not satisfies none of those.
What does properly identified evidence look like? Goli, Reiley and Zhang ran an 18-month field experiment involving more than seven million Pandora users, experimentally varying advertising load, and found that a personalised reallocation policy could raise subscription profits by 7% without reducing advertising profits (13). That is unusually strong. It is also one platform, one intervention and a finite horizon, and the authors also estimate a reduction in consumer welfare, which is a useful reminder that company profit and customer benefit do not always move together.
Three separate editorial ideas are worth keeping distinct throughout any CLV discussion:
- Observed CLV: the value associated with customers who received or engaged with personalisation.
- Predicted CLV: what a model forecasts under assumed conditions.
- Incremental CLV: the difference between expected value with the intervention and without it, after incremental costs.
Only the third supports a claim that personalisation caused anything.
The data question, and where I have landed on it
There is a widespread assumption that more behavioural data produces better personalisation and therefore better lifetime value. My experience makes me more cautious about that than the market generally is.
Behavioural data can be useful. But its meaning is often less stable than teams assume. A browse, a click or a dwell event can represent several different intentions, and the resulting model has to be validated against an outcome rather than accepted because the segment looks plausible. Two customers with near-identical browsing histories are frequently doing entirely different things.
Purchase data behaves differently. It is unambiguous, comparatively small, and describes something the customer actually did. In my own work, transactional and deliberately declared data have produced decisions that were easier to explain and easier to evaluate.
Herm is built around purchase data, so I have a commercial interest in that interpretation. Which is why I would treat it as an operating thesis to test rather than a settled conclusion. The test is straightforward enough: hold behavioural and purchase-based segmentation against the same outcome under the same design, and see which one survives.
What I would separate from that is the value of asking. One of the largest consumer-electronics companies in the world would not commit to a targeted discount on a product group until it had more than behavioural inference to work from. We ran a custom survey collecting declared data directly from end users, and the conversion contribution from respondents was strong enough that the brand had the confidence to discount.
The case illustrates a narrower point than it might appear to. Even a very large organisation may prefer deliberately declared information when the commercial decision is consequential and behavioural inference is ambiguous. Declared data is not only an SME workaround for lacking scale.
The costs a CLV calculation must include
Most marketing CLV calculations subtract product margin and stop. That produces a number that is reliably too high.
Net revenue should be after discounts and promotional credits, expected refunds and cancellations, expected returns, and loyalty rewards attributable to the transaction. IFRS 15 treats variable consideration such as discounts, rebates, refunds and incentives as part of the transaction price, and for goods sold with a right of return, revenue is not recognised for products expected to be returned (14).
Incremental customer costs should include, where material:
- cost of goods, content or service provision;
- fulfilment, distribution and payment costs;
- service and support costs;
- returns processing and lost product value;
- retention campaigns and promotional subsidies;
- personalisation technology, data, modelling and delivery costs;
- channel or partner commissions;
- acquisition cost, where the question is total relationship economics.
The Institute of Management Accountants’ CLV work computes lifetime value from contribution margin, where contribution already incorporates service cost, and then subtracts acquisition cost (15). ACCA and KPMG’s profitability guidance goes further, recommending end-to-end analysis covering the whole value chain rather than product margin alone (16).
One judgement call worth stating explicitly. For an incremental decision about a personalisation programme, do not load the model with fixed costs that will not change because of it. For a fully allocated customer-profitability report, disclose the overhead allocation method rather than silently excluding it. These are different exercises and conflating them produces arguments that cannot be resolved.
When personalisation reduces customer value
Four mechanisms, all of which I have seen produce good-looking reporting:
Discounting people who would have bought anyway. The most expensive failure in this field, and the least visible, because every engagement metric improves while margin quietly leaves.
Acceleration mistaken for growth. Reminders that move purchases forward without increasing annual volume.
Unprofitable cross-selling. Adding categories to customers whose service cost, returns or promotion dependency exceeds the additional margin (4).
Training customers to wait. Persistent discount-led personalisation teaches people that the normal price is the wrong price, which reduces full-price purchasing across the entire base rather than just the treated segment.
Building a business case that survives scrutiny
Five things, in this order.
State the driver. Which of order value, frequency or lifespan is this meant to move, and by what mechanism?
State the counterfactual. What comparison establishes the effect, and is it a holdout, a phased rollout or a before-and-after? Say which.
Use contribution, not revenue. After incentives, delivery, service and returns.
Separate realised from forecast. A measured effect on a treated population and a projection onto the full eligible population are different figures. Label them separately, and state the assumption that the effect persists at scale rather than burying it inside the multiplication.
Attach uncertainty. An interval and a sensitivity test on the retention and discount-rate assumptions. A single point estimate for a modelled future value implies a confidence nobody has.
For how to construct the comparison itself, see our guide to measuring personalisation effectiveness.
Frequently Asked Questions
Does a 5% increase in retention really increase profits by 25% to 95%?
Not as usually stated. The claim traces to Reichheld and Sasser's 1990 Harvard Business Review article, which appears to have given the range as 25% to 85%; Bain was publishing 25% to 95% by 2000. More importantly, peer-reviewed analysis by Pfeifer and Farris shows the increases apply to the discounted value of customer cash flows, not to company profit, and that the exercise adds roughly five percentage points to the retention rate rather than improving it by 5% relative. It is a comparative accounting model holding margins and retention spending fixed, which real businesses cannot do.
How do you prove personalisation increased customer lifetime value?
You need random assignment or another defensible causal design, analysis by assigned treatment rather than by who engaged, contribution margin rather than revenue, all treatment costs included, and an observation window long enough for the behaviour to appear. Comparing customers who used personalisation against those who did not proves nothing, because more engaged customers are more likely both to use the feature and to generate future value anyway. Netflix states the underlying point plainly: personalisation models are purely associative.
Does cross-selling increase customer lifetime value?
Sometimes, and it can also reduce it. Research across five company databases in consumer and business markets found that between 10% and 35% of customers who cross-bought were unprofitable, and those customers accounted for 39% to 88% of total customer losses. The mechanisms included limited total spending despite buying across categories, revenue reversals, excessive service requests and promotion-driven purchasing. Cross-sell rate is not a proxy for profitability, so incremental product revenue has to be reduced by cost of goods, discounts, returns and service before it counts.
Are long-term customers always more profitable?
No. Reinartz and Kumar analysed three years of daily customer data from a catalogue retailer and concluded that long-life customers are not necessarily profitable customers, with tenure and profitability related only imperfectly. The accurate proposition is narrower: retention creates value when the discounted incremental contribution from the relationship exceeds the incremental cost of retaining and serving it. Retaining a customer whose expected future margin is negative destroys value regardless of how long they stay.
Which costs should a CLV calculation include?
Net revenue after discounts, expected refunds, returns and attributable loyalty rewards, then incremental customer costs including cost of goods, fulfilment and payment, service and support, returns processing, retention campaigns and subsidies, personalisation technology and delivery, partner commissions, and acquisition cost where the question is total relationship economics. Most marketing CLV calculations subtract product margin and stop, which reliably produces a figure that is too high. For an incremental decision, exclude fixed costs that will not change; for a fully allocated report, disclose the allocation method.
References
- Gupta, S., Hanssens, D., Hardie, B., Kahn, W., Kumar, V., Lin, N., Ravishanker, N. and Sriram, S. Modeling Customer Lifetime Value. Journal of Service Research, 9(2), 139-155, November 2006. https://journals.sagepub.com/doi/10.1177/1094670506293810
- Fader, P. S., Hardie, B. G. S. and Shang, J. Customer-Base Analysis in a Discrete-Time Noncontractual Setting. Marketing Science, 2010. https://pubsonline.informs.org/doi/10.1287/mksc.1100.0580
- Ascarza, E. and Hardie, B. G. S. A Joint Model of Usage and Churn in Contractual Settings. Marketing Science, 2013. https://pubsonline.informs.org/doi/10.1287/mksc.2013.0786
- Shah, D., Kumar, V., Qu, Y. and Chen, S. Unprofitable Cross-Buying: Evidence from Consumer and Business Markets. Journal of Marketing, 2012. https://journals.sagepub.com/doi/10.1509/jm.10.0445
- Munz, K. P., Jung, M. H. and Alter, A. L. Name Similarity Encourages Generosity: A Field Experiment in Email Personalization. Marketing Science, 10 April 2020. https://pubsonline.informs.org/doi/10.1287/mksc.2019.1220
- Reichheld, F. F. and Sasser, W. E. Zero Defections: Quality Comes to Services. Harvard Business Review, September-October 1990. https://hbr.org/1990/09/zero-defections-quality-comes-to-services
- Reichheld, F. F. and Schefter, P. E-Loyalty: Your Secret Weapon on the Web. Bain & Company, 1 July 2000. https://www.bain.com/insights/e-loyalty-your-secret-weapon-on-the-web/
- Reichheld, F. F. Loyalty Rules!, Chapter 1. Harvard Business School Press, 2001. https://www.bain.com/contentassets/29f74ec417fa4e36a1d7d7e7479badc5/loyalty_rules_chapter_one.pdf
- Pfeifer, P. E. and Farris, P. W. The Elasticity of Customer Value to Retention: The Duration of a Customer Relationship. Journal of Interactive Marketing, May 2004. https://journals.sagepub.com/doi/10.1002/dir.20006
- Reinartz, W. J. and Kumar, V. On the Profitability of Long-Life Customers in a Noncontractual Setting. Journal of Marketing, October 2000. https://journals.sagepub.com/doi/abs/10.1509/jmkg.64.4.17.18077
- Ascarza, E. Retention Futility: Targeting High-Risk Customers Might Be Ineffective. Journal of Marketing Research, 55(1), 80-98, 2018. https://journals.sagepub.com/doi/10.1509/jmr.16.0163
- Netflix Technology Blog. A Survey of Causal Inference Applications at Netflix. 21 May 2022. https://netflixtechblog.com/a-survey-of-causal-inference-applications-at-netflix-b62d25175e6f
- Goli, A., Reiley, D. H. and Zhang, H. Personalizing Ad Load to Optimize Subscription and Ad Revenues. Marketing Science, 2024. https://pubsonline.informs.org/doi/10.1287/mksc.2022.0357
- IFRS Foundation. IFRS 15 Revenue from Contracts with Customers. https://www.ifrs.org/content/dam/ifrs/publications/pdf-standards/english/2021/issued/part-a/ifrs-15-revenue-from-contracts-with-customers.pdf
- Bonacchi, M. and Perego, P. Measuring and Managing Customer Lifetime Value: A CLV Scorecard and Cohort Analysis in a Subscription-Based Enterprise. Management Accounting Quarterly, Institute of Management Accountants, Fall 2012. https://www.imanet.org/~/media/IMA/Files/Home/Insights-and-Trends/MAQ/MAQ-2012/MAQ_-Fall-2012/qfall2012_bonacchi-pdf.pdf
- ACCA and KPMG. Profitability and Cost Analysis: An Eye on Value. April 2016. https://www.accaglobal.com/content/dam/ACCA_Global/Technical/fin/Profitability-and-cost-analysis-An-Eye-on-Value.pdf
Use this guide as a source
If it settled an argument in your reporting, cite it — and add Herm as a preferred source so the next answer you get from search or a model comes from work with its definitions attached.
