Can the same discount budget generate more revenue?
Could a US home decor retailer earn more by distributing the same discount budget differently? In the reported 11 days of its trial, the new coupon selection method raised average revenue per user by 4.5% compared with the existing method, using the same budget. How discounts are allocated can matter alongside how much a business spends on them.
ACADEMIC WORK REVIEWED
Personalized Promotions in Practice: Dynamic Allocation and Reference Effects
ResearchersJackie Baek; Will Ma; Dmitry Mitrofanov
arXiv · 29 December 2025 · v1, not peer reviewed · Preprint · Source language: English
English review: 24 October 2026 · The source date appears in the citation above.
Editorial review draft
The source has been checked; the review still carries its original editorial status. DijitalPi did not conduct a customer experiment for this paper.
LET'S READ THE RESEARCH TOGETHER
We first explain the researchers' question, method and findings. We then discuss how to interpret the results, clearly separating DijitalPi's commentary from the source.
01 / WHAT DID THE RESEARCHERS WANT TO UNDERSTAND?
Can selecting coupons using customer history increase revenue within the same discount budget?
The researchers seek to determine which discount each customer should receive in daily emails while respecting the overall promotional budget. They also examine how customers using their best recent discount as a comparison point could shape future offer sequences. This comparison mechanism is called a reference effect.
Explaining the context · DijitalPi commentary
DijitalPi commentary: Assessing a discount solely by today’s advertised percentage may leave out relevant information. Offers the same customer received recently deserve attention when examining responses to a new coupon. The study provides a reason to consider both the allocation of the budget across customers and their offer history.
02 / HOW WAS THE RESEARCH CONDUCTED?
How did they test the question?
The study took place at an unnamed large US online home decor retailer. Model development used 90 days of historical data from 360,000 customers and 61 features. In May-June 2024, 20 million customers were randomly split 50/50 between a new coupon algorithm and the existing algorithm, which had been used for several years. The authors received only aggregate, relative results for an 11-day period of the test. A separate group of 150,000 customers receiving a random coupon each day for 11 days provided data for the reference-effect analysis.
Understanding the method · DijitalPi commentary
The existing algorithm grouped customers largely by tenure and average spending. The new method used a model that learned from past data to predict the chance of a purchase to select a 10%, 12%, 15%, 17% or 20% coupon. It did not use generative AI. A company employee adjusted the budget parameter daily to match the existing algorithm’s promotional budget. The comparison therefore did not pit personalisation against sending everyone the same offer.
The deployed algorithm aimed to increase expected next-day revenue. Attempts to predict spending amounts failed, so the model focused on purchase probability and did not account for the cost of goods sold. The main outcome was average revenue per user, with contribution margin reported separately. The model was not retrained during the reported 11-day period. The outcomes extended beyond email opens or clicks.
03 / RESEARCH FINDINGS
The new coupon allocation generated more revenue at the same budget during the reported short period
The researchers reported a 4.5% increase in average revenue per user against the existing algorithm. The increase was statistically significant in a one-sided test: p < 0.01. The reported increase in total revenue over 11 days was also 4.5% for the equally sized groups. Contribution margin increased by 5.6%. The source gives no separate significance result for this outcome. The authors suggested a shift towards higher-margin products as a possible explanation, but did not test it. In the separate random-coupon data, a higher best recent discount was associated with a lower propensity to purchase and was more predictive than the average of past discounts. These associations were small in magnitude. More than 99.9% of customer-day observations contained no purchase.
Average revenue per user
4.5%Increase for the new algorithm against the existing algorithm at the same promotional budget. Restricted to the reported 11-day period, with p < 0.01 in a one-sided test.
Contribution margin
5.6%Reported increase in the new algorithm group. A change in product choice was not tested as the explanation. This should not be read as an increase in net profit.
DijitalPi commentary: Consider testing an alternative to your existing coupon grouping that incorporates customer behaviour and offer history. Evaluate revenue per user and contribution margin at the same promotional budget. Investigate which allocation works in your own data rather than adopting the paper’s effect size as your target.
The reference-effect association is not a field estimate of the long-term harm from repeatedly sending large discounts. Under its assumptions of one customer, no budget and no external changes, the theoretical model supports a cycle of weak discounts throughout the memory period followed by one strong discount. That cycle was never field-tested. The deployed algorithm targeted the next day alone. Results from daily US retail emails cannot be directly extended to Türkiye, Turkish-language campaigns or other sectors.
05 / CONCLUSION AND OPEN QUESTIONS
What did we learn, and what do we still not know?
The findings concern one US home decor retailer and do not establish long-term gains or outcomes for Turkish-language campaigns in Türkiye. This preprint provides short-term randomised field evidence that a more detailed allocation method can improve an existing coupon algorithm at the same promotional budget. Absolute revenue amounts and results for the entire test were not shared, nor were the most recent data available to the authors. The algorithm became the default in August-September 2024, but the company deprioritised personalisation at the end of 2024, switched to sending everyone the largest discount every day and phased out the algorithm by early 2025. This prevented testing the theoretical cycle. The authors were not company employees, but they developed the algorithm and helped prepare production code. The company supplied aggregate, relative results. The source contains no funding or competing-interest statement, and the paper has not undergone peer review.
Our open question at DijitalPi: if your coupon decisions considered each customer’s highest recent discount, which outcome would you test within the same budget? The illustrative scenarios are application proposals; DijitalPi conducted no coupon experiment for this review.
We created these scenarios to make the topic concrete. They are not cases from the paper or measured client results.
EXAMPLE 01
E-commerce: allocating coupons within the same budget
An illustrative clothing retailer groups customers by purchasing history and sends everyone in each group the same coupon.
Would selecting coupons individually generate more revenue than the existing grouping?
Open the recommendation for example 01+
Keep the current grouping as the comparison and test the new allocation at an equal promotional budget.
Randomly assign customers to compare the existing method with the new coupon selection. Record coupon history, revenue per user and contribution margin for each group. Set the test duration and decision criterion before starting. The paper’s short-term result does not imply the same increase for clothing or Turkish-language emails.
EXAMPLE 02
Tourism: a new discount follows a stronger offer
An illustrative hotel is preparing a booking email with a smaller discount for people who previously received a larger offer.
Does response to the new offer vary with the discount a customer recently received?
Open the recommendation for example 02+
Record previous offers and assess responses to the new discount alongside that history.
Link each customer’s offer dates and discount levels with realised booking revenue. Keep stay dates and offer conditions consistent in the comparison. Do not substitute email opens for booking revenue. The paper’s association with previous discounts cannot, on its own, explain why hotel customers decline a smaller offer.
EXAMPLE 03
Education: treating a discount cycle as a ready-made schedule
An illustrative language school is considering several small-discount offers followed by a large-discount offer.
Can we directly adopt the discount cycle described in the paper?
Open the recommendation for example 03+
Treat the cycle as a separate proposal to test, rather than a field-validated prescription.
Retain the existing offer schedule as the comparison group. Track the sequence of offers, enrolment revenue per user and discount costs during the test. Assess performance beyond the large-discount day alone. The paper’s cycling result is theoretical and supplies no tested schedule for education services.
TRY IT WITH YOUR TEAM
Include coupon history in your measurement
Link discounts to purchase records by customer and date, then design an equal-budget test that retains the existing allocation method as the comparison. Contribution margin is revenue minus costs that vary with sales. Evaluate revenue per user and contribution margin separately, without treating a short-term result as a long-term gain. This is a proposal that DijitalPi has not implemented.
The academic work belongs to the researchers named above. This page contains DijitalPi's explanatory review and original business examples; it is not a full translation of the paper.
Cookie preferences