Predictive Email Tests: Keep Creative Changes Clear

If a predictive segment gets a different email, offer, and sending schedule from the comparison group, a better result does not tell you which change helped. Before producing variants, decide whether you are testing audience selection or the message itself.

Migma is useful at the creative boundary: turn the agreed question into a reviewed email, keep the approved version identifiable, and record which deliberate change each variant contains. The audience and measurement owners then implement the experiment in the system that actually selects recipients and records outcomes.

This guide is published by Migma. The workflow below is an editorial recommendation for teams evaluating predictive targeting, not a claim that Migma creates predictive models or assigns holdout groups. Product guidance was checked on October 20, 2026.

Write the question in one sentence

Start with a business question that names the decision. For example: “Does this targeting policy improve repeat purchases compared with our existing approach?” That differs from asking whether a shorter email works better for the same eligible audience.

Brevo’s predictive-segmentation article discusses control groups and the importance of outcomes beyond campaign volume. Use that as a reason to examine your test design, rather than assuming a model score proves a customer needs a particular message.

Have the measurement owner define the comparison, outcome, observation window, and requirements for a useful result. Do this before asking Migma for new creative. A polished variant can otherwise become a reason to run a test that cannot answer the original question.

Separate an audience test from a creative test

In an audience-selection test, keep the message sufficiently stable that the comparison can examine the targeting decision. In a creative test, hold the relevant audience rules stable while varying the agreed element of the message. A broader experiment may vary both, but it needs a design and analysis that account for those differences.

Consider a fictional outdoor store introducing a care guide after a purchase. One team wants to compare a predictive selection policy with its existing eligibility rules. Another wants to compare a short explanation with a detailed guide. Combining those requests in one casual split would make the result difficult to interpret.

Write the two decisions separately. The first needs a reviewed message that can be used consistently across the planned audience comparison. The second needs clearly defined versions and an assignment method the analyst accepts. Neither question requires claiming the model knows a recipient’s intentions.

Prepare the stable message in Migma

Give Migma approved facts about the product, care task, and destination. Include the audience description as context for the writing, while keeping actual eligibility with the audience owner. Use reviewed brand settings and references to keep the identity recognizable.

For the care-guide example, specify the exact product family, the one maintenance action the guide explains, and the approved destination. Avoid inserting a discount merely because a score suggests someone might lapse. That would introduce another material change and a different customer promise.

Review the opening, button, and practical instructions. A reader should understand what they can do next without being told that they are “high risk” or have been assigned to an internal predictive category. Keep the content grounded in useful product information.

Save the accepted version and record its purpose. If a later change is necessary, make the measurement owner aware before the test continues. A safety or accuracy correction may be essential, but it still changes what was delivered.

Keep a small version record

A version record can be a short document beside the approved creative. It should identify the message, date, source facts, intended test, changed elements, and reviewer. Give the exported file or destination template an unambiguous label.

For a creative comparison, describe the difference precisely. “Version B moves the care instruction above the product image” is more useful than “Version B is more engaging.” Preserve the agreed offer, destination, subject, or other elements that are outside the question being tested.

Check that the final destination matches the approved version. Migma’s export documentation explains that another platform can add wrappers, fill variables, and apply sending rules. Compare the actual template after handoff; the original preview alone cannot establish what recipients will receive.

Review the holdout implementation with its owner

The audience owner should explain how people enter the comparison, what the holdout receives, and whether other messages can reach them during the observation window. Write down the intended behavior before activation.

Do not choose an arbitrary holdout percentage from a blog example. Suitable allocation and duration depend on the outcome, available audience, expected effect, and analysis method. Ask the analyst to establish those choices and the limits of the eventual conclusion.

Also check ordinary operational constraints. Permission, suppression, recent purchases, and required service messages can affect eligibility. A marketing experiment should not prevent a necessary account or order notice. Have the relevant owner resolve these boundaries using the actual sending system.

Read results without rewriting the history

When the observation window ends, compare the recorded assignment and delivered versions with the original plan. Note material deviations before interpreting the outcome. Missing data, changed offers, or overlapping campaigns may limit what the team can conclude.

Migma’s results guide describes campaign metrics and side-by-side A/B results. It also states that Migma does not choose an A/B winner automatically. These product reports do not replace an analyst’s design for a separate predictive-targeting or incrementality experiment.

Use the result to make the next brief more specific. If a finding is uncertain, preserve that uncertainty. If a change earns further investigation, explain what will stay fixed in the next test. Start with one reviewed Migma email and one answerable question; create more variants when the design calls for them.

Frequently Asked Questions