Polish E-Commerce Group — IP Warming at Scale Across 6 Sending Domains and 4 European ISPs

A Warsaw-based e-commerce group operating six distinct retail brands undertook a complete email-infrastructure rebuild following an acquisition that brought all brands under a single operational team. Each brand had previously operated independently, with its own ESP account, its own sending domain, its own historical reputation, and its own marketing team running campaigns to its own subscriber base. The consolidation required migrating all six brands to shared dedicated infrastructure while maintaining completely isolated reputation profiles between them — a failure in one brand's reputation must not affect any other under any operational condition. The group could not accept the structural risk that had been the cost of the prior shared-ESP arrangements where brand-level reputation was effectively co-mingled.

Total production volume across the six brands was 28 million messages per month at steady state, with significant skew during seasonal peaks (Black Friday in late November, Christmas in mid-December, post-holiday returns and clearance in January). The infrastructure required 16 new sending IPs, all starting with zero reputation history. European ISP coverage was the operational priority: GMX, Web.de, and T-Online (Germany), Orange and OVH (France) collectively accounted for 38% of the group's recipient base across all brands, and these were precisely the ISPs that respond least forgivingly to volume increases on freshly-warmed infrastructure. The warming program had to succeed across multiple new IPs, multiple new domains, multiple ISPs with different warming tolerances, and a calendar that included Black Friday in Week 7 of the planned ramp.

The IP assignment was designed to group IPs by ISP destination rather than by brand. This is the architectural decision that distinguishes this engagement from a more conventional brand-by-brand allocation, and the rationale is operational. Building IP reputation with a specific ISP cluster — Gmail and Yahoo together, the German national ISPs together, the French national ISPs together — produces stronger per-ISP reputation signals in a shorter timeframe than spreading volume across all destinations from each IP. A Gmail-dedicated pool sending exclusively to Gmail recipients across all six brands accumulates Gmail-specific reputation faster than a brand-dedicated pool sending to all destinations from each IP. The brand isolation is preserved through DKIM signing identity (each brand has its own DKIM key) and DNS-level domain separation, not through IP-level partitioning.

European ISPs — particularly the German national providers GMX, Web.de, and T-Online — have measurably different warming behaviour than Google and Microsoft. They respond more conservatively to volume increases, maintain elevated deferral rates for longer during the early warming phase, and apply per-IP connection limits that punish over-connection patterns more aggressively than the US-dominant providers. The warming pace at the German pool was deliberately slower than at the Gmail pool: where the Gmail pool ramped from 80 messages per IP per hour in Week 1 to 600 per IP per hour by Week 7, the German pool ramped from 40 per IP per hour to 250. The difference compounds across the warming period, but the slower pace produced lower deferral rates throughout warming and a more stable reputation outcome at production volume.

Week 7 of the warming schedule coincided with Black Friday — a 340% volume spike requirement that the warming infrastructure could not absorb at the cadence required. Rather than attempting to send full Black Friday volume through the partially-warmed IPs, which would have produced a reputation event that no subsequent operational discipline could fully recover, we implemented a hybrid approach: Black Friday campaigns sent through a combination of the warming infrastructure (up to warming-appropriate limits) supplemented by a temporary cloud relay for overflow volume. The cloud relay was a separate sending path with its own reputation history, contracted specifically for the seasonal overflow window and decommissioned after Black Friday concluded. This maintained warming progress on the dedicated infrastructure without sending above the reputation-safe threshold, while ensuring that the brand teams' Black Friday campaigns reached their full intended audience through a path whose reputation could absorb the burst.

Technical Assessment: Infrastructure Layers Examined

The post-warming assessment examined three layers that distinguish multi-domain warming from single-domain warming: the per-brand DKIM and DMARC architecture that enforced reputation isolation despite shared per-ISP IP pools, the orchestration layer that coordinated warming pace across 16 IPs and 4 ISP clusters simultaneously, and the integration with the seasonal-volume cloud-relay overflow that kept Black Friday from compromising the warming program.

Per-Brand DKIM and DMARC Architecture

Each of the six brand domains has its own DKIM signing key under a per-brand selector — brand-a-2024._domainkey.brand-a.pl, brand-b-2024._domainkey.brand-b.pl, and so on. When PowerMTA signs an outbound message, the signature uses the brand-specific key matched to the From-domain, regardless of which IP pool the message is routing through. A message from Brand A to a Gmail recipient signs with Brand A's DKIM key and routes through the Gmail-pool IPs; a message from Brand B to a Gmail recipient signs with Brand B's DKIM key and routes through the same Gmail-pool IPs. Receiving ISPs evaluate each brand's authentication identity independently in their reputation calculations, even when the underlying sending IP is shared. This is the technical basis for the IP-by-ISP-destination architecture: reputation isolation between brands is enforced at the authentication layer, not at the IP layer, because authentication-layer isolation is sufficient and IP-layer isolation would degrade the warming outcome.

Multi-IP Warming Orchestration

Coordinating the warming of 16 IPs simultaneously across 4 ISP clusters required orchestration logic that conventional MailWizz-style campaign sending does not provide natively. The platform's PowerMTA configuration includes per-pool, per-ISP throttle parameters that scale weekly across the warming period, with adjustment logic triggered by deferral-rate measurements at each ISP cluster. Week 1 starts every pool at conservative volume; Week 2 reviews actual deferral rates per ISP cluster and either advances volume (if deferral rate is below 3%) or holds the current configuration (if above 3%). The decision is per-cluster, not per-IP — all four IPs in the German pool advance or hold together based on the German cluster's aggregate deferral rate, ensuring that pool-level reputation builds uniformly rather than developing per-IP variance that would produce inconsistent delivery experiences for recipients on the same ISP cluster.

Cloud-Relay Overflow Integration

The Black Friday cloud-relay overflow was integrated through a separate listener on the platform's PowerMTA infrastructure, with credentials specific to the campaign-management tooling and a routing decision that diverted overflow volume to the contracted cloud relay rather than to the warming pools. Once Black Friday concluded, the listener was decommissioned and the relay contract closed; the routing decision reverted to standard pool assignment. The integration was operationally clean because PowerMTA's routing flexibility allowed the temporary detour without modifying the brand-team-facing tooling — the campaign managers continued to submit campaigns through the standard interface, with the overflow logic invisible at their layer. This kind of seasonal-overflow architecture is reusable for any future seasonal peak that exceeds the dedicated infrastructure's intended capacity envelope.

Infrastructure Rebuild: Configuration Decisions

Per-pool warming pace tuned to ISP-cluster behaviour. The Gmail/Yahoo pool warmed at the most aggressive cadence (Gmail/Yahoo tolerate volume increases more permissively than European ISPs), the Microsoft pool at intermediate cadence, the German national pool at the most conservative pace (German ISPs require longer reputation-stabilization windows). Each pool's warming schedule was independent — Gmail pool advancing to Week 4 thresholds did not trigger automatic advance for the German pool. The independence is critical because applying Gmail-appropriate cadence to German ISPs would have produced deferral spikes that compounded into reputation events; applying German-appropriate cadence to Gmail would have unnecessarily delayed the Gmail warming completion.

Per-brand suppression registries with cross-brand awareness. Each of the six brands maintains its own suppression registry tracking unsubscribes, complaints, hard bounces, and ISP feedback-loop signals for that brand. Cross-brand awareness is added through a registry-overlay layer: a recipient who unsubscribes from Brand A's communications is flagged in a shared overlay that other brands' campaign builders can optionally consult. This is operationally useful because some recipients may have unsubscribed from one brand specifically (legitimate per-brand opt-out) while remaining engaged with another brand in the group; other recipients have unsubscribed from any communication using their email address regardless of which brand sent it. The overlay surfaces both patterns without forcing a one-size-fits-all suppression policy across brands.

FBL processing routed by brand DKIM identity. Complaint feedback loops are subscribed at every major ISP that supports them. When a complaint arrives, the FBL processor identifies the brand from the DKIM signing identity in the original message and routes the complaint to that brand's suppression registry — not to a unified group-wide suppression. This preserves the per-brand reputation attribution: a complaint on Brand A's traffic affects Brand A's reputation registry without contaminating the suppression status of recipients who happen to be active on other brands' lists. The brand-level FBL routing is the operational complement to the brand-level DKIM signing, completing the per-brand reputation attribution that the architecture is built around.

Operational Monitoring: What Changed Permanently

Per-pool deliverability dashboards reviewed daily, per-brand reports weekly. The infrastructure operations team reviews per-ISP-pool delivery rates, deferral rates, and complaint rates daily, with attention focused on any pool where metrics deviate from established baselines. The brand campaign managers receive per-brand weekly reports covering their specific brand's metrics across all ISP clusters. Treating these as different cadences for different audiences — operational team daily on infrastructure metrics, brand teams weekly on brand metrics — surfaces the right information to the right team at the right frequency. Pool-level operational issues are caught quickly by the team responsible; brand-level patterns are surfaced to brand teams in time to inform their next campaign decisions.

Per-pool warming-pace adjustment as ongoing capability. The warming infrastructure was not decommissioned after the initial 10-week warming completed. The same orchestration logic remains operational and is reused whenever new IPs are added to existing pools (capacity expansion as the group grows) or new pools are created (new ISP clusters added if the recipient base shifts geographically). The first re-warming event happened nine months post-deployment when 4 additional IPs were added to the Gmail pool to support volume growth; the warming completed in 6 weeks because the existing pool reputation could be inherited progressively rather than building from zero, and the orchestration logic that coordinated the original 16-IP warming generalized cleanly to the additive case.

Quarterly brand-portfolio reputation review. Across the six brands, the operations team reviews quarterly which brands are operating within healthy reputation parameters and whether any brand is at risk of producing pool-wide reputation degradation through unmanaged behaviour. The shared-per-ISP-pool architecture works only if all brands maintain consistent operational standards; the quarterly review enforces that consistency by identifying outliers early. In the post-engagement year, two brands were flagged for elevated complaint-rate trends and provided targeted consultation on list-quality and content; both brands recovered their metrics within a quarter. The review process is what keeps the per-ISP-pool architecture viable across an evolving brand portfolio.

Multi-domain warming at scale rewards architectural decisions that conventional single-domain warming does not need to make. The IP-by-ISP-destination allocation produces stronger per-ISP reputation in shorter timeframes than the brand-by-brand allocation that operational teams default to, but it requires per-brand reputation isolation to be enforced at the authentication layer rather than the IP layer. This trade-off works only if all brands sharing each ISP pool maintain consistent sending standards; it is appropriate for post-acquisition consolidation where unified operational standards have been established, and it is inappropriate for multi-tenant agency models where per-tenant isolation must be absolute regardless of behaviour. The architectural decision is conditional on the operational context, not categorical.

Warming programs that overlap with seasonal volume peaks require an explicit overflow strategy from the start, not an emergency response when the peak arrives. The hybrid cloud-relay approach used here for Black Friday is the operational pattern for any warming program where calendar constraints prevent the warming from completing before peak season — temporary contracted overflow capacity preserves the warming reputation by absorbing the volume that the warming infrastructure cannot safely send. The cost of the seasonal overflow is small relative to the cost of a reputation event that compromises the warming program; treating it as an integral component of the warming plan rather than an emergency contingency is the discipline that makes warming-during-peak-season operationally sustainable.