A constructed case study: a deliberately dirty 35-row dataset audited under extreme constraints, because sometimes n=34 is all you get.
Orders July 1–29, 2026 · 34 observed orders after dedupe; order 1029 known missing · prepared by Alex Kwon for the Growth Manager · every figure independently recomputed
Social produced $102 of contribution for the month, 1.8% of the total from 21% of orders: 4 of its 7 orders were returned, and 70% of the revenue it booked came back ($451 of $644 gross), where every other channel sits at 13% or lower. Return rates are floors—the extract records no return dates, so recent orders may yet come back.
| Channel | Orders | Returns | Return rate | Net revenue | Contribution |
|---|---|---|---|---|---|
| Organic | 9 | 0 | 0% | $2,621.10 | $2,472.10 |
| Paid Search | 9 | 1 | 11% | $1,788.16 | $1,667.16 |
| 9 | 0 | 0% | $1,474.30 | $1,381.55 | |
| Social | 7 | 4 | 57% | $192.87 | $101.87 |
Monitor orders at ≥20% discount averaged $43 contribution per order against $407 for monitors at ≤15%; two of the four deep-discount orders returned. With four orders we cannot separate discount depth from channel—2 of the 4 are North/Social—so this is reported as a description, not a cause.
Buckets cover all 9 monitor orders; none fall between 15% and 20%. One order (1025) sold at 90% off, grossing $32.90 against $27.00 shipping—mis-entry or real clearance needs the order record.
Watch item: order 1016 is 10 mouse units in a file where no other order exceeds 4 of anything—42% of July mouse units in one order. Likely a bulk buyer; the alternative risk scenario (inventory drain followed by mass return) would show up as these units coming back, which is why their post-export status is requested. Excluded from per-order averages it would dominate.
Five planted defects, five found (the FINDINGS.md scorecard, Q1–Q4). Three were corrected at load; two are flagged by order id and left in place, because the data cannot tell a typo from a real clearance or a bulk buyer from a drain. The pair in rows 1 and 2 is consistent with a single paste-over during entry or export: totals were simultaneously overstated by one known order and understated by one unknown one.
| Defect | Action | Revenue impact |
|---|---|---|
| Order 1028 entered twice (Q1) | deduplicated | −$151.98 gross |
| Order 1029 absent from extract (Q1) | requested from source | unknown; revenue understated |
| Region “west” on order 1024 (Q2) | normalized to West | none (classification only) |
| Discounts of 50% and 90% on orders 1009, 1025 (Q3) | flagged, not corrected; records requested | none until adjudicated |
| Order 1016: 10 mouse units, 42% of mouse volume (Q4) | flagged; sensitivity run with and without it | none; drives per-order averages |
⚠ limits
Whether Social’s return problem is the channel, its promotions, or its products: the deep-discount monitors and the Social returns overlap, and n=4–7 cannot separate them. Any trend: this is one month with no baseline. True return rates: no return dates in the extract, so every rate shown is a floor.
| Request | Settles |
|---|---|
| Return reasons per returned order | defective/open-box vs changed-mind |
| Promotion & ad detail for Social, July | what those 7 orders were acquired with |
| Order 1029 from the source system | the missing record |
| Order records for 1009, 1025 | whether 50% and 90% discounts were authorized |
| Return dates & policy terms | real return rates instead of floors |
| Order history, prior 12 months | baseline, seasonality, whether any of this is new |
| Post-export status of order 1016; C115 history | bulk buyer vs drain pattern |
Definitions: revenue = units × unit price × (1 − discount%); returned orders count $0 revenue with shipping treated as sunk; contribution = net revenue − shipping, manufacturing cost out of scope per the brief. Basis: 34 orders after the corrections above. Analysis: Alex Kwon. Full workings, findings with row ids, and runnable SQL: the ecommerce-case-study repository (FINDINGS.md, sql/00–04).
Cite: Kwon, A. (2026). July Orders Review: an e-commerce data-quality case study. github.com/collapseindex/ecommerce-case-study (machine-readable: CITATION.cff in the repository). License: report CC BY 4.0; code Apache-2.0. Disclosure: AI-assisted implementation; methodology and validation framework developed in prior research; every figure recomputed from raw data by the runnable code in the repository.
Alex Kwon · ask@collapseindex.org · more case studies · github.com/collapseindex