Direct answer
How should a marketplace handle refund liability?
Most destination-charge marketplaces absorb refunds mechanically — the rails leave no choice — and recover from sellers deliberately: reversing transfers inside payout windows, netting slower items against future earnings, and writing off whatever neither reaches. The whole model rests on three prerequisites: written seller terms granting recovery rights, communication templates making notices routine, and monitoring fast enough that recovery starts while balances still exist.
The ladder
Reverse now (funds reachable, relationship tolerates immediacy) → net later (active seller, smaller stakes, terms allow setoff) → write off (dormant account, sub-floor amount, EV negative). Ordered by urgency matching payout physics rather than org-chart convenience. Every rung needs the figure computed identically — proportional expectation minus actual reversals — or rungs argue with each other.
Terms and tone prerequisites
Setoff and recovery clauses in seller agreements convert collections into administration. Notice templates convert debits into dialogue. Neither requires legalese heroics — they require existing before the first recovery attempt, which is when leverage and goodwill are both highest.
Monitoring makes the ladder operational
Without continuous detection the ladder exists theoretically while month-end delivers aged findings whose bottom rungs already expired. Real-time findings enter at the correct rung automatically — which is the difference between having a policy and running one.
Common questions
Is there industry-standard practice?
Convergent practice, rarely documented publicly — this page is essentially that documentation. Variation lives in floors and tone, not structure.
Small marketplaces too?
The ladder scales down intact; floors rise relative to volume. Ten careful refunds weekly justify the same discipline as ten thousand.