How LTV.ai finds the hour each customer actually reads email
"Send at 10am Tuesday" is folklore. LTV.ai predicts each customer's best hour from their own behavior, measured on opens and clicks together, with clicks anchoring the signal so privacy-inflated opens cannot skew it, and puts that prediction to work twice.
Nairit BanerjeeHead of Machine Learning
Ask when to send an email campaign and you will get a confident answer from every marketing blog ever written: Tuesday at 10am. Thursday at 2pm. The answer is always a single hour, for the entire list, borrowed from someone else's aggregate study.
But a list is not a person. It contains commuters who read email at 7am, parents who read it at 9:30pm after bedtime, and night owls at 1am. One send time is guaranteed to be wrong for most of them.
How it is done today, and why it is weak
Global send-time rules average away the individual. Tools that do offer send-time optimization typically learn it from opens, and opens have quietly become one of the most polluted signals in email.
Apple's Mail Privacy Protection pre-fetches messages and fires open events whether or not a human ever looked, at times that reflect Apple's servers, not the reader. A send-time model trained on opens is, for a large share of any list, learning the schedule of a datacenter.
The result: confident per-user send times that are wrong in a systematic, invisible way.
Why we think this is worth getting right
Timing gates everything downstream. An email that arrives while the customer is asleep is buried by morning, whatever its subject line says.
And because we run live subject-line experiments during sends, we have a second, sharper reason to care: experiments need clean feedback fast, and feedback only arrives when recipients are actually looking. Getting send time right per customer improves every send and makes the platform's own learning loops faster and more trustworthy.
How LTV.ai approaches it
An Optimal Send Time model scores every customer on each hour of the day, a 24-point engagement curve built from their own history of when opens and clicks have actually occurred, and predicts the single hour they are most likely to engage. Not a coarse morning or afternoon bucket: their hour.
The model reads both signals, but not equally. A click cannot be pre-fetched by a privacy proxy, it is a human decision, so clicks anchor the model while opens are weighted with appropriate suspicion where Apple Mail Privacy Protection is likely inflating them. The result is a curve that reflects when a customer reads and acts, not when a server touched the message. Like every model in the prediction suite, it is validated out of sample and reports insufficient data rather than guessing on thin history.
At send time, the campaign's audience is scheduled so each customer's email lands at their predicted hour: the same campaign, arriving at thousands of individually right moments.
The second job: powering honest experiments. The same per-customer windows shape the batches of the live subject-line optimization. When a bandit send is split into staggered batches, customers are allocated to the batches that overlap their predicted best hours, and the small early explore batches are filled first, with the closest hour matches.
Those batches carry the experiment: the more of their recipients are genuinely awake and reading, the more opens arrive quickly, and the cleaner the signal. It is confound control. An arm's poor performance means a weak subject line, not an unlucky send hour.
How it stays honest and compounds
The model retrains automatically as behavior accumulates, so a customer whose routine shifts, a new job, a new timezone, new habits, drifts to a new window within days rather than quarters.
And because hourly predictions feed both delivery and experimentation, every improvement in send-time accuracy compounds: emails land at better moments, experiments read cleaner, and the subject lines those experiments crown are more trustworthy.
Frequently asked questions
Doesn't Apple's privacy protection break send-time optimization? It breaks purely open-based optimization. Ours uses opens and clicks together, anchored on clicks, a signal privacy proxies cannot fake.
What about customers with no click history? The model reports insufficient data for them and they are scheduled normally. No fabricated windows.
Part of the machine learning behind LTV.ai.
See it on your store: book a demo.