Talk to an expert
Most email campaigns go out at a single scheduled time — say, 10 AM on Tuesday — chosen because it tested well on average across the whole list. The problem is that "on average" hides enormous variation in when individual subscribers actually check their inbox.
A subscriber who works night shifts might not look at email until 4 PM. A parent doing school drop-off may not open anything before 9:30 AM. Someone in a different time zone receives that 10 AM send at 7 AM local time, before they're awake.
Most opens happen within the first hour or two of delivery. After that, newer messages push it further down the inbox. By the time the subscriber checks in, the email is effectively invisible.
Batch sending also creates deliverability problems at scale. Sending hundreds of thousands of messages simultaneously generates traffic spikes that mailbox providers flag. Sudden volume bursts can trigger throttling or spam filtering, which means some subscribers may not receive the message in their primary inbox at all.
Batch sending treats a subscriber list as one audience with one optimal send window. In practice, a list of 100,000 subscribers includes people across multiple time zones, work schedules, and email-checking habits. A single send time can only align with a small subset of those habits. The rest receive the message when they are unlikely to see it.
Per-recipient send time optimization assigns each subscriber a delivery window based on their individual engagement history — typically the hour or two when they most frequently open or click. The system analyzes past open timestamps, builds a per-user activity profile, and schedules each message accordingly.
The results hold across industries. An analysis of over 5,000 campaigns from 90 clients found that per-recipient timing lifted conversions by at least 15% across every messaging channel, with some channels exceeding 30% 1.
Individual results break down further. KFC Ecuador saw a 15% increase in open rates after switching to individualized send times 1. OneRoof, a property listings platform, recorded a 23% increase in click-to-open rates and a 57% uplift in unique clicks 1. foodora measured a 9% increase in email click-through rate and a 26% reduction in unsubscribe rate 1.
When messages arrive at times subscribers are already checking email, those messages feel less intrusive — fewer subscribers unsubscribe.
Instead of competing with every other marketing email sent at the same popular hour, each message arrives when fewer emails are in the inbox and the subscriber is already reading email. The message appears near the top of the inbox rather than buried under other mail.
Send time models operate on behavioral signals that each subscriber generates through normal email interaction. The primary inputs are open timestamps, click times, and the gaps between delivery and engagement. Some implementations extend this to website browsing sessions and purchase activity, showing when each person is active across channels.
The model's job is to convert these raw timestamps into a per-user engagement profile. According to a peer-reviewed framework published in MDPI, the key data features include historical engagement times, day-of-week patterns, and recency-weighted interaction signals 2. More sophisticated implementations calculate the best sending hour (0-23) for each day of the week, producing separate optimizations for opens versus clicks — generating as many as 16 prediction fields per profile 3. This distinction matters because the hour someone opens an email often differs from the hour they click through to take action.
Time zone data is a required input. Without it, a model cannot distinguish between a subscriber who engages at 9 AM Eastern and one who engages at 9 AM Pacific. Most platforms infer time zones from IP geolocation at signup, then refine that estimate as behavioral data accumulates.
The cold-start problem is the most practical challenge. A new subscriber has no engagement history, so the model has nothing to learn from. Production systems require at least one month of delivery and tracking data before individual predictions become reliable 3. Until that threshold is met, new subscribers typically receive messages based on aggregate patterns — the best-performing times across similar audience segments. As each subscriber interacts with more messages, their individual profile gains accuracy and the model replaces segment-level estimates with personalized predictions.
Each new open or click updates the profile, which means send time predictions adapt as subscriber behavior changes — after a new job, a time zone change, or a different daily routine.
Every send time model faces a technical tradeoff: should it keep sending at the time that has performed best so far, or should it test alternative windows that might perform better? This is the multi-armed bandit problem, a framework from decision science 4.
The two competing strategies are exploitation and exploration. Exploitation chooses the best option based on current knowledge — sending at the time slot with the highest observed engagement. Exploration tries new options that may yield better outcomes, accepting lower short-term engagement for better long-term performance 4.
A model that only exploits will lock onto a single send time and never adapt. A model that only explores will keep testing random windows and never capitalize on what it has learned. Effective send time optimization requires both, running simultaneously.
Subscriber behavior changes over time — optimal send times shift as people change jobs, move time zones, or adjust routines 5. A time slot that drove strong engagement six months ago may now land during a commute. Without testing alternative windows, the model cannot detect these shifts.
Most implementations handle this automatically. The system allocates most sends to the highest-performing time slots while directing a controlled fraction to alternative windows. When an alternative outperforms the current best, the model shifts sends to the new time. This rebalancing runs continuously without manual intervention.
Send time optimization extends beyond email to any channel where timing affects engagement. Push notifications, SMS messages, and even paid ad impressions all benefit from per-user delivery timing — but each channel has different constraints.
SMS messages show higher open rates than email but arrive as phone alerts, making poorly timed messages harder to ignore. A message arriving at 11 PM may get read, but it damages brand perception and increases opt-outs. Effective STO for SMS must weigh not just predicted engagement but also acceptable delivery windows, which vary by region and regulation. Several platforms now apply ML-based timing predictions to both email and push channels 6, and some extend optimization across email, SMS, and push within a 24-hour window 7.
Push notifications present a different challenge: they compete for attention on a locked screen alongside other alerts. ML models trained on per-user push interaction history can predict the windows when a specific user is most likely to engage 8. Ad delivery timing follows a similar principle — serving an impression when a user is actively browsing yields higher click-through than delivery during inactive hours.
Cross-channel STO also enables coordination. Without a unified timing layer, a user might receive an email, a push notification, and an SMS within minutes of each other. A shared optimization model staggers delivery across channels, reducing fatigue and keeping response rates higher for each message.
Send time optimization requires two foundational components: a behavioral data layer that captures user engagement signals, and integration with the platform that actually delivers messages.
The behavioral data layer typically involves placing a tracking tag on your website or app, or connecting to an existing CDP that already collects open times, click timestamps, and session activity. This data feeds the ML models that build individual engagement profiles. Most platforms require at least one month of historical delivery and tracking data before predictions become reliable 3.
Integration with your ESP or marketing automation platform determines how optimized send times actually reach recipients. Some ESPs offer built-in send time features that require toggling a single setting, while standalone ML platforms need API-based connections or native integrations with your existing stack. Built-in ESP features are faster to activate but typically use simpler models trained only on that platform's data. Standalone ML platforms can ingest behavioral signals from multiple channels—web, app, email, push—producing per-user predictions based on a broader set of behavioral signals, but require more setup work to connect data sources and delivery endpoints.
Typical implementation timelines range from days for built-in ESP features to several weeks for standalone platforms that require a consulting engagement 3. Timelines depend on data infrastructure readiness and the number of integrations needed.
List size affects prediction accuracy. Models need sufficient engagement data per recipient to identify individual timing preferences. Lists under a few thousand active recipients may not generate enough signal for per-user optimization, defaulting instead to segment-level or cohort-based timing. Larger lists with consistent engagement history produce the most granular predictions.
Any optimization system needs a measurement framework that isolates its actual impact from noise. Holdout testing provides this by splitting your audience into two groups: a treatment group that receives ML-optimized send times and a control group that receives messages at a fixed time. The control group acts as a baseline, showing what performance looks like without optimization.
For group sizing, allocate 5-20% of your traffic to the control group. Too small and you lack statistical power; too large and you sacrifice potential gains across your audience. Each group needs at least 1,000 recipients to reliably detect meaningful differences in performance.
Do not evaluate results early. The optimization model needs time to learn individual behavioral patterns, and short windows amplify the effect of seasonal variation. Run the test for a minimum of 90 days before drawing conclusions.
When comparing groups, look beyond open rates. Measure click-through rates, conversion rates, and revenue per recipient. The core calculation: subtract control group conversions from treatment group conversions, then divide by control group conversions. This gives you the incremental lift directly attributable to send time optimization, excluding gains that would have occurred without optimization.
Send time optimization assumes the message itself is sound and that it can reach the inbox. When either condition fails, better timing produces negligible results.
Low open rates often trace back to content problems: weak subject lines, irrelevant offers, or generic messaging that gives recipients no reason to open. If subscribers consistently ignore emails regardless of when they arrive, the content needs work before timing adjustments will matter.
List quality is another common root cause. Stale lists filled with inactive addresses, role-based accounts, or contacts who never opted in generate low engagement by definition. No algorithm can make a disinterested recipient open an email. Regular list hygiene — removing hard bounces, sunsetting long-term non-openers, and validating new addresses — is a prerequisite for any optimization work.
Deliverability failures precede both content and timing as a factor. Sender reputation, authentication protocols, sending cadence, list quality, and engagement history all determine whether messages reach the inbox at all 9. Spam filters also evaluate structural signals like subject line patterns and image-to-text ratios 9. If emails land in spam, optimizing send times is irrelevant.
The diagnostic question: are engaged subscribers seeing your emails and choosing not to act? If yes, timing is a reasonable variable to test. If no, fix the upstream issue first.
Q: Do I need a data engineering team to set up send time optimization?
A: No. Most STO platforms are designed for marketing teams to configure through a dashboard. The technical requirements are limited to installing a tracking tag on your website and connecting your email service provider or messaging platform through a native integration or API.
Q: How much historical data do I need before STO starts working?
A: Most models require roughly one month of engagement data to generate initial predictions for a subscriber. Accuracy improves as more behavioral signals accumulate. New subscribers with no history are handled through cold-start methods that fall back on population-level patterns until individual data is available.
Q: How long does it take to see measurable results?
A: Plan for a multi-month holdout test before drawing reliable conclusions. Early directional signals may appear within weeks, but subscriber behavior fluctuates enough that shorter windows produce noisy data. Statistical confidence requires adequate sample sizes in both the optimized and control groups.
Q: Is STO worth the cost for small lists?
A: The lift from per-subscriber timing is real, but small lists make it harder to reach statistical significance and may not generate enough revenue gain to justify the subscription. Lists under a few thousand active subscribers often benefit more from fixing segmentation, content, or deliverability first. The break-even calculation depends on your average revenue per email and the vendor's pricing tier.
Q: Does STO keep working if a subscriber's behavior changes?
A: Yes. Modern STO systems use continuous learning, updating each subscriber's optimal window as new engagement data arrives. The models balance exploiting the current best time against exploring alternative windows to detect behavioral shifts.
Q: Can STO compensate for poor email content or list hygiene?
A: No. Timing optimization affects when a message arrives, not whether the content is relevant or the recipient is engaged. Sending a weak offer at the perfect time still produces a weak result. STO delivers incremental lift on top of fundamentals that are already working.
Q: How does Almeta ML handle send time optimization?
A: Almeta ML analyzes each subscriber's historical engagement data to calculate a per-user optimal send window. The platform connects to email services, push notification systems, and advertising platforms via API, so the computed send times translate directly into delivery scheduling across channels. Configuration runs through a visual dashboard with no coding required.
Predict Customer Behavior with Almeta ML
Real-time actionable predictive metrics for your website