Which customers need an incentive to return?
At an on-demand grocery marketplace, I helped turn broad, expensive reactivation into segmented offers and a repeatable testing process.
- My role
- Analysis, testing & operational support
- Original tools
- SQL · Excel · BI
- Deliverable
- Segmented campaign strategy
The problem
Sending the same coupon broadly made reactivation costly. Which customers responded, and which offer could justify its subsidy?
What I did
- Compared offer and holdout outcomes across prior incentive sensitivity and basket value.
- Tested coupon terms and merchant funding, then supported customer tags, offer matching, and recurring readouts.
Incentive sensitivity helped identify where an offer added more response.
The moderately sensitive group showed the largest offer-versus-holdout gap. A high sensitivity label alone was insufficient to choose the best audience.
Compare response with the segment’s own baseline
View data & definitions
The offered group is 1.5 times its holdout in every sensitivity band. Return-rate gaps are descriptive comparisons. Original allocation and uncertainty are unavailable; the replacement counts do not establish statistical significance or a randomized causal effect.
| Sensitivity | Holdout users | Holdout returners | Offered users | Offered returners | Rate gap (pp) |
|---|---|---|---|---|---|
| Low | 40,000 | 440 | 60,000 | 720 | 0.10 |
| Light | 60,000 | 780 | 90,000 | 1,620 | 0.50 |
| Moderate | 50,000 | 300 | 75,000 | 1,050 | 0.80 |
| High | 50,000 | 350 | 75,000 | 675 | 0.20 |
Decision: Use response differences alongside customer economics to choose who receives an incentive.
Lower barriers and merchant funding created different cost options.
For lower-basket-value customers, the lower-threshold offer generated more estimated additional returns than the higher-threshold offer. Merchant funding reduced platform cost, with a slightly different observed response.
Read response and incremental platform cost together
View data & definitions
Within each basket-value audience, three strategies share one holdout. Baseline returns and costs are scaled to the offered group’s size. Merchant funding reduces platform spend but is not assumed to remove all merchant or business costs. Similar observed response rates do not prove equivalence. The sensitivity exhibit uses a separate population.
| Key | Audience | Offer | Holdout rate | Offer rate | Estimated extra / 100k | Platform CU / estimated extra returner |
|---|---|---|---|---|---|---|
| 1 | Lower basket value | Lower threshold | 0.90% | 1.70% | 800 | 8.38 |
| 2 | Lower basket value | Higher threshold | 0.90% | 1.50% | 600 | 12 |
| 3 | Lower basket value | Merchant-funded | 0.90% | 1.65% | 750 | 5.3 |
| 4 | Higher basket value | Lower threshold | 1.70% | 2.10% | 400 | 54.5 |
| 5 | Higher basket value | Higher threshold | 1.70% | 2.30% | 600 | 48 |
| 6 | Higher basket value | Merchant-funded | 1.70% | 2.28% | 580 | 33.45 |
- Lower basket value · lower threshold
- Lower basket value · higher threshold
- Lower basket value · merchant-funded
- Higher basket value · lower threshold
- Higher basket value · higher threshold
- Higher basket value · merchant-funded
Decision: Match coupon terms to the audience and evaluate merchant-funded options on both response and platform cost.
Some customers would have returned without the offer.
The control-adjusted platform cost is 8,040 CU: 8.38 CU per estimated additional returner.
This calculation makes the comparison and its assumptions inspectable. It also keeps natural reactivation out of the campaign’s estimated added response.
Build a reusable offer strategy around response and cost.
What I delivered: Segmentation and offer evaluation, supported by recurring audience selection and result queries. The historical decks document the progression into customer tags and merchant-offer matching used in operations.
Comparison assumptions, cost definitions & source notes
Based on my JD Daojia quarterly work reports and operations-sharing material. Original work used SQL, spreadsheets, BI reporting, and campaign evaluation. Python charts and the replacement counts are newly produced for this portfolio.
- The reconstruction defines dormant users as previous purchasers with no order in the prior 30 days and uses a seven-day return horizon. This does not assert the original test used that exact eligibility or horizon.
- Estimated extra returns = offered returns − holdout returns × offered users ÷ holdout users. The same size adjustment is applied to total platform subsidy costs before dividing by estimated extra returns.
- These comparisons assume comparable groups. The original assignment mechanism, balance checks, confidence intervals, and statistical power are not available here.
- Sensitivity and basket-value examples are different populations. Parallel offers share one holdout per basket-value audience; their return estimates must not be summed as one campaign.
- Costs are platform subsidy, including background subsidy in each arm; they exclude acquisition infrastructure, merchant contribution, and full business profit.
The historical qualitative rollout is distinct from the reconstructed estimates. Original decks, customer IDs, offer systems, screenshots, and SQL remain private.
Download comparison data