Real Estate Platform — Property Alert and Agent Communication Email Infrastructure

An Australian real estate platform headquartered in Sydney serves the ANZ residential property market through a search engine, agent contact directory, and saved-search alert system. The platform's email volume is dominated by two streams: property alerts triggered when a new listing matches a saved search (1.2 million daily during normal market conditions, peaking to 1.8 million during seasonally active periods), and agent-to-buyer communications routed through the platform's contact intermediation product. Buyer engagement with property alerts had been declining for 18 months — from 22% click-through to 8% — without the operations or product team having a unified theory of why.

By mid-2024, deliverability degradation had become impossible to ignore. Outlook inbox placement was at 69%, Gmail at 76%, with continuing weekly decline. Complaint rates had climbed from a historical 0.04% to 0.12% — still below Gmail's 0.3% threshold, but trending in a direction that would cross it within a quarter. The marketing team had been treating the click-rate decline as a creative problem (better subject lines, better property photography, better email design); the deliverability team had been treating the inbox placement decline as an authentication or sending-volume problem; neither had identified the structural cause that linked both metrics.

The engagement reframed the problem before scoping the technical work. Property alert email is not a marketing channel where volume optimization compounds; it is an operational notification channel where relevance compounds. A buyer who receives one alert per week that genuinely matches their search will engage at materially higher rates than the same buyer receiving daily alerts where 60% are tangential matches. Sending less of the right thing was both the deliverability fix and the product improvement, and treating them as the same intervention was the engagement's central principle.

1. Weeks 1–3: Search-lifecycle modelling and audience segmentation

   Built a search-state classifier from product event data: searches where the user has visited a property listing in the past 14 days are classified "active"; searches where last visit was 14–60 days are "dormant-recoverable"; searches with no visit in 60+ days are "dormant-suppressed". The 4.1 million subscriber base segmented to 760,000 active searchers, 540,000 dormant-recoverable, and 2.8 million dormant-suppressed. Alert sending was immediately stopped for the dormant-suppressed segment (no cold-cutoff complaints emerged because these users were not engaging anyway), and dormant-recoverable users were moved to a re-engagement track.

2. Weeks 3–5: Relevance threshold tightening and per-search calibration

   The match threshold for sending an alert was raised from 40% to 70% criteria match, with criteria weighted by buyer-defined importance (price match weighted higher than property type match, suburb match weighted higher than bedroom count). Per-search relevance models were trained on each user's prior listing-view behaviour: a buyer who consistently viewed listings in three specific suburbs would not receive alerts for similar listings in adjacent suburbs unless they had also viewed properties there. The threshold change reduced alert volume from 1.2 million to roughly 760,000 daily — but the alerts that did send were materially more likely to match the buyer's actual search behaviour.

3. Weeks 5–8: Dedicated infrastructure provisioning and warming

   Provisioned 6 dedicated IPs in Sydney for primary buyer-alert traffic (geo-routing trust pattern with major Australian ISPs) and 2 dedicated IPs in Singapore for the platform's growing international-buyer segment (Mainland China, Singapore, Malaysia residents purchasing Australian property). Authentication baseline rebuilt: SPF cleaned to active services only, 2048-bit DKIM with rotation cadence, DMARC progressed from p=quarantine to p=reject across the warming period. Active-searcher traffic moved to dedicated infrastructure first; dormant-recoverable traffic remained on the legacy infrastructure during their re-engagement window.

4. Weeks 8–12: Re-engagement campaign, full cutover, agent stream migration

   Three-stage re-engagement campaign ran to the 540,000 dormant-recoverable subscribers across six weeks. The first email asked whether the search was still relevant; the second offered to update search criteria; the third was an explicit opt-out invitation. 142,000 confirmed continued interest and were merged into the active-searcher segment; 87,000 explicitly opted out; the remaining 311,000 did not respond and were moved to suppressed. Agent-to-buyer communications, a separate stream with different sending patterns, migrated to its own pair of dedicated IPs to isolate B2B agent traffic from B2C buyer alerts.

Technical Assessment: Infrastructure Layers Examined

Engagement Filtering and the Volume Penalty

Gmail and Outlook both apply engagement-based reputation scoring at the domain level. A domain where 80% of recent recipients do not interact with mail receives lower placement than a domain where 40% interact, even if absolute complaint rates are similar. The platform's 4.1 million subscriber base meant that every send to dormant subscribers was contributing a no-engagement signal that aggregated across the domain's reputation calculation. Suppressing the 3.28 million dormant-suppressed segment was not just a list-hygiene improvement; it was a direct intervention on the engagement-rate denominator that drives the reputation score.

The arithmetic is straightforward. Pre-suppression, 1.2 million daily sends produced approximately 96,000 daily clicks (the 8% click rate). Post-suppression, 760,000 daily sends produced approximately 167,000 daily clicks (the new 22% click rate). Absolute engagement nearly doubled even though volume dropped 37%, because the volume reduction concentrated sending on subscribers who were operationally interested. The reputation signal at every major mailbox provider responded within four weeks: Gmail moved from "Low" to "High" domain reputation, Outlook moved from "Bad" to "Good" sender reputation, and inbox placement climbed in both environments.

Relevance Matching as a Reputation Lever

The 40% to 70% threshold change deserves examination because it is an unusual deliverability intervention — most engagements treat relevance matching as a product concern unrelated to email infrastructure. In this case the connection was direct. An alert at 45% relevance is statistically more likely to be ignored than acted on; an ignored alert produces a no-engagement signal; cumulative no-engagement signals depress reputation. By raising the relevance bar to 70%, the platform reduced send volume but increased the per-send engagement rate, producing a positive shift in the reputation signal even if absolute engagement counts had stayed flat.

The per-search calibration was the operationally subtle component. A naive 70% threshold would over-restrict for buyers with broad searches (looking for properties in multiple suburbs across a wide price range) and under-restrict for buyers with narrow searches (specific suburb, narrow price band). Weighting the criteria by buyer-defined importance — and by inferred importance from listing-view behaviour — produced a relevance model that adapted to each search's actual scope. The model was simpler than a full machine-learning system; it was rules-based with weights, transparent and auditable, and it was sufficient for this use case.

Australian and APAC ISP Behaviour

The Australian mailbox-provider environment differs from European and North American patterns in operational ways relevant to the migration. Telstra (the dominant national ISP) and Bigpond accept email at lower per-IP throughput rates than Gmail or Outlook but tolerate longer connection persistence — sending behaviour optimized for Gmail's high-throughput pattern produces deferrals at Telstra. The Sydney-origin dedicated IPs were calibrated for the throughput and persistence patterns Australian ISPs prefer, with per-domain configuration explicit for the Telstra and Bigpond domains. The Singapore-origin IPs were calibrated for the international-buyer segment whose recipients used a mix of Gmail (dominant in Mainland China and Malaysia consumer segments), Outlook, and regional ISPs.

Infrastructure Rebuild: Configuration Decisions

Search-state as a sending precondition. The PowerMTA injection layer queries the search-state classifier before accepting an alert message into the queue. An alert message generated for a search whose state has shifted to "dormant-suppressed" since the last classifier update is rejected at injection rather than queued and sent. This redundancy with the campaign-builder filter prevents a class of errors where a search state transitions during the brief interval between alert generation and send.

Subdomain split for buyer alerts and agent communications. Buyer alerts send from alerts.platform.com.au; agent-to-buyer communications send from contact.platform.com.au; transactional account email from account.platform.com.au. The split allows independent reputation development for each stream. Buyer alerts have higher volume but lower per-message engagement; agent communications have lower volume but higher per-message engagement and higher business value (every agent-buyer interaction is a potential transaction). Mixing the two on a single domain would have averaged the metrics rather than letting each stream optimize for its actual characteristics.

Search-criteria audit logging. Every alert send records the search criteria, the listing criteria, and the relevance score that triggered the send. This audit trail enables post-hoc analysis of the relevance model's accuracy: when a recipient unsubscribes after an alert, the logs show what relevance score the alert received, allowing identification of model edge cases where the score was high but the relevance was perceived as low. Three model adjustments were made in the first six months post-deployment based on this audit, all targeting specific edge cases (luxury properties at the upper end of search ranges, properties in school-catchment-adjacent suburbs that buyers had not explicitly searched).

Operational Monitoring: What Changed Permanently

Engagement rate as the leading deliverability metric. The operations team treats per-send engagement rate (clicks divided by sends, measured at 7-day rolling window) as the leading indicator of deliverability health rather than the lagging inbox-placement rate. Engagement decline precedes inbox-placement decline by 2–4 weeks at every major mailbox provider with engagement-based filtering. Monitoring the leading indicator means corrective action — usually relevance-model adjustment or search-lifecycle threshold tuning — can happen before the placement metric reflects the underlying degradation. Two such corrections were made in the first year post-deployment, both surfaced by engagement-rate drift before placement metrics moved.

Quarterly relevance-model audit. The relevance model is reviewed quarterly against actual buyer behaviour: did high-relevance-score alerts produce higher click rates? Did unsubscribers correlate with low-relevance alerts they received in the prior 14 days? Are there listing types or suburbs where the model systematically over- or under-scored relevance? The audit produces incremental adjustments rather than wholesale rewrites; after six quarters, the model accuracy has improved measurably without requiring a major redesign.

Search-state distribution monitoring. The active / dormant-recoverable / dormant-suppressed distribution is monitored monthly. Healthy distribution is roughly 60% active, 20% dormant-recoverable, 20% dormant-suppressed. Drift outside these bounds indicates a structural issue: a sustained decline in the active percentage suggests the platform is acquiring users without converting them to active searchers; a sustained climb in dormant-suppressed suggests the lifecycle thresholds may be too aggressive. The monitoring exists not just to maintain deliverability but to surface product-funnel issues that would otherwise be invisible until they had already affected business metrics.

Property alert email is operational notification, not marketing communication. Buyers who save searches expect alerts to match those searches with reasonable specificity; they tolerate occasional broad matches but interpret a sustained pattern of low-relevance alerts as either a platform that does not understand their search or a platform that is using their inbox to push inventory rather than support their goal. Both interpretations produce disengagement, complaint, and ultimately unsubscribe — and each of those signals depresses the deliverability that allows the platform to reach the buyers whose searches are still active.

The case for treating relevance and deliverability as a single problem is general beyond real estate. Any platform sending notification email triggered by user-defined criteria (job alerts, price-drop alerts, content recommendations, dating matches) faces the same dynamic: low-relevance notifications produce disengagement, disengagement produces reputation damage, reputation damage produces broader delivery degradation that affects even the high-relevance notifications that would have engaged. Treating relevance as a deliverability lever — and the matching algorithm as a deliverability tool — is unusual practice but produces compounding benefits in any context where notification volume is determined by a matching threshold the platform controls.