Skip to content

SYSTEM Cited by 1 source

Zalando Marketing Services (ZMS)

What it is

Zalando Marketing Services (ZMS) is Zalando's retail-media / sponsored-products advertising business — a two-sided marketplace in which advertisers (brands) compete for ad placements, customers see the resulting sponsored products at the digital point of sale, and the whole system is bounded by a finite shared resource: advertisers' campaign budgets. Auctions allocate placements; ML models (pCTR predicted click-through-rate and pCVR predicted conversion-rate, retrained daily on the last 7 / 28 days) rank and price bids.

ZMS is distinct from Zalando's general-purpose Octopus experimentation platform: ZMS runs its own experimentation framework tailored to the constraints of an adtech marketplace, where the standard user-level A/B test breaks (Source: sources/2026-08-31-zalando-scaling-reliable-experimentation-in-a-two-sided-adtech-marketplace).

The experimentation problem

Because advertisers compete through a finite shared budget, a classical user-split A/B test violates SUTVA: Treatment and Control variants drawing from the same campaign wallet interfere via Cannibalization Bias — a more efficient variant wins more auctions, drains the shared budget, and starves the other variant, biasing the measured effect.

What ZMS built (2023–2025)

  • Budget Split — partition each campaign's budget into per-variant "sub-campaigns" (proportional to the traffic split) so Treatment and Control auction against isolated budgets. The instance of per-variant-resource-isolation.
  • Orthogonal Concurrency — split campaign budgets into 2ⁿ orthogonal buckets (1A2A / 1A2B / 1B2A / 1B2B for two experiments) so multiple experiments run concurrently without cross-experiment cannibalization.
  • Residual-interference monitoring — even with Budget Split, ZMS watches four residual SUTVA-violation channels: algorithmic campaign steering on combined metrics, pre-experiment ML training-data contamination, manual advertising-ops interventions, and timing / right-censoring effects.
  • Causal-inference fallback — "when A/B testing is not possible, or if we cannot guarantee isolation, we pivot to causal inference methodologies (like non-experimental counterfactual analysis)" — the same design stance as Octopus's quasi-experimental tooling.

Impact

ZMS experimentation scaled from 8 experiments in 2023 (no Budget Split) to 60+ experiments in 2025 with the isolation properties above — turning a flaky, sequential process into a reliable, concurrent one and making Budget Split "a prerequisite for trust" for PMs, analysts, and applied scientists.

Seen in

Last updated · 766 distilled / 2,225 read