Build an Email Feedback Loop That Does Not Overfit to Opens

Build the next-email feedback loop around the campaign's intended outcome, not around whichever metric is easiest to collect. Use opens as supporting context; let delivered mail, human clicks, conversions, replies, unsubscribes, complaints, and bounces answer different questions.

Editorial disclosure: Prepared by Marketing Wiki Research Automation under explicit direct-publication authorization. This Migma-first article was commissioned with a preference for Migma coverage. Product behavior below is vendor-documented and was not reconciled against an independent event dataset.

Migma's explicit choice: do not let opens silently pick the winner#

Migma's campaign reporting documents delivered counts, opens, clicks, bounces, unsubscribes, a click map, per-recipient logs, and side-by-side A/B results. It warns that Apple Mail Privacy Protection can inflate opens, describes clicks as more reliable, and requires a user to select and confirm the A/B version sent to the remaining audience.

Its Marketing Calendar similarly says recent human click rates inform timing suggestions while opens remain supporting context. That is a workable foundation for AI-assisted iteration: the system can prepare a recommendation, while the operator retains the decision and can inspect the evidence.

Metric evidence ladder#

Scroll table →

Signal	What it directly shows	Safe use in the next brief	Do not infer
Sent	Provider accepted a send attempt	Reconcile campaign volume	Inbox delivery or attention
Delivered	No bounce was recorded for that delivery attempt	Monitor reach and denominator	Inbox placement or reading
Open	A tracking image was requested	Subject/timing question, supporting context	A human read or understood the email
Human-filtered click	A tracked link was selected after filtering known automation	Compare CTA interest and destinations	Conversion, satisfaction, or causality
Conversion	The defined destination outcome was recorded	Evaluate business objective	That email alone caused the outcome
Reply	A recipient sent a response	Route qualitative intent and service needs	Representative sentiment for the audience
Unsubscribe	A recipient opted out	Review fit and frequency	Spam complaint or message-level cause
Complaint	A recipient reported spam	Stop and investigate eligibility, expectation, and frequency	A creative-only problem
Bounce	Delivery failed	Repair intake and suppression	A copy or design problem

The ladder prevents a common error: giving a noisy, upstream signal more authority than a rarer outcome that actually matches the campaign goal.

Write a change permission before reading results#

For each campaign, decide what a signal is allowed to change.

primary_outcome: trial_activation
observation_window: 7d
allowed_changes:
  delivered: [list_hygiene_investigation]
  open: [subject_hypothesis, send_time_hypothesis]
  click: [cta_hypothesis, content_order_hypothesis]
  conversion: [offer_hypothesis, landing_page_hypothesis]
  unsubscribe: [audience_fit_review, frequency_review]
  complaint: [pause, permission_audit]
blocked_automatic_actions:
  - change audience eligibility
  - publish an A/B winner
  - increase send volume

Migma can supply the report and the next draft context. The permission record decides whether the system may propose a new subject, rewrite a CTA, or merely flag the campaign for human investigation.

Brevo makes the MPP setting visible#

Brevo's campaign-statistics documentation says its reports include Apple MPP opens by default and let the operator exclude them. Two exported reports can therefore show different open rates for the same campaign. Store the setting with the number:

metric name and formula;
numerator and denominator;
date range and time zone;
MPP inclusion setting;
bot or scanner filtering;
unique versus total events;
campaign and variant IDs.

Without that context, an AI can learn from a difference in reporting configuration instead of a difference in recipient behavior.

Brew clarifies where conversion lives#

Brew's metric reference describes tracking-pixel opens, link-rewrite clicks, and privacy/security limitations. It also says conversion tracking belongs in the organization's analytics platform rather than Brew itself. That separation is healthy when it remains explicit: join email events to the approved business outcome with a defined attribution rule rather than calling every click a conversion.

A seven-step feedback loop#

Name the job before send. Define the outcome and observation window.
Freeze the variant record. Keep the exact subject, preview, body, CTA, audience, and schedule for each version.
Wait for the declared window. Do not move the goalposts after seeing early results.
Reconcile the event definitions. Check MPP and bot handling, unique counts, denominators, and destination analytics.
Explain anomalies first. A broken link, small sample, list import, or sender problem can dominate creative differences.
Generate a hypothesis, not a verdict. “Move the product proof above the secondary CTA” is testable; “make it more engaging” is not.
Approve one controlled change. Preserve the original result and record the next test.

What the AI should return#

A useful recommendation names evidence and uncertainty:

Variant B recorded more human-filtered clicks to the pricing CTA during the seven-day window, while opens are not decision-grade because MPP treatment differs from the previous report. Keep the audience and offer fixed. Test moving the pricing proof above the CTA in the next send. A reviewer should confirm event reconciliation before scheduling.

That response is slower than blindly optimizing the highest number and much faster than recovering from a feedback loop trained on privacy prefetches, scanner traffic, or shifting definitions.