Reviews
Notes on email segmentation from the field
Email segmentation for 2027 covers purpose, eligibility, data lineage, identity, time rules, overlap, permissions, fairness, validation, experiments, and incidents.
What to take away
- A segment is a reproducible rule that changes a named decision for an eligible population. If no decision changes, do not build it.
- Write membership, exclusion, precedence, time, unknown-value, refresh and retirement rules so another analyst can reproduce them.
- Eligibility and suppression come before targeting. A missing value never counts as a favorable answer.
- Version every production segment, test edge records and overlaps, and dry-run each destination before release.
- The 2027 shift is documentation, not law: definitions, owners and change history now travel with the rule.
Name the decision a segment changes
An email segment is a reproducible rule that carves part of an eligible audience so a stated message changes: sent, held, sequenced, localized or prioritized. Without that change, a saved filter is only a reporting preference. Segments sit under a wider email marketing strategy that decides which choices the program may make.
Does the segment change a decision?
Does it change which approved message is sent?
Record need, channel, message class, harm
Complexity without clear purpose
Documentation is the difference from prior practice, more than any new rule. Definitions now carry an owner, a purpose, a version and a change history instead of living in one analyst's head.
Separate audience, segment, cohort and persona
The four words get used as synonyms and should not be. An audience is everyone eligible for a communication. A segment is a rule selecting part of it. A cohort shares an event or starting period and exists for analysis. A persona is a research archetype with no contact-level status.
Keeping them apart stops a fictional persona from quietly becoming a targeting rule, and stops a measurement cohort from being activated as a send list.
Audience concepts compared
Audience
- Definition
- Eligible group
- Scope
- Whole group
- Risk
- Too broad
Segment
- Definition
- Reproducible rule
- Scope
- Part of audience
- Risk
- Hidden targeting
Cohort
- Definition
- Shared event
- Scope
- Starting period
- Risk
- Analysis only
Persona
- Definition
- Research archetype
- Scope
- Not contact fact
- Risk
- Fictional rules
Example: one complete segment rule
Eligible: confirmed address, market US or CA, permission granted, evaluated in the account time zone at send. Include: last completed order between 1 and 180 calendar days before evaluation, a typical win-back window. Exclude: refunded-only records, open complaints, central suppression, employee and test records, and anyone contacted more than 3 times in 7 days. Unknown market or permission: excluded. Precedence: safety and service messages outrank this segment. Refresh: nightly, with suppression and permission rechecked at send. Retire when the replacement path ships.
Two layers sit behind that paragraph. The sentence is what a reviewer approves. The builder logic underneath must return the same population under the same version number. A rule this shape feeds email campaign planning, which decides the offer and the order of sends.
Eligibility, suppressions and precedence
| Layer | Question it answers | Typical contents |
|---|---|---|
| Permission | May we message this person at all? | Consent, objection, unsubscribe, complaint |
| Eligibility | Is the contact in scope now? | Market, language, verified address, active relationship, age or sector limits |
| Targeting | Which message does the contact get? | Lifecycle stage, recent order, engagement, product status |
| Precedence | Who wins when rules overlap? | Safety, service, contractual, lifecycle, promotional |
| Suppression | Who must never receive it? | Legal hold, incident hold, invalid address, employee and test records |
Precedence belongs in a visible table, not in tribal knowledge. Frequency limits belong there too, because two segments can each look reasonable while together over-contacting one person.
Central suppression beats local campaign filters once more than one tool can send. Exclusions need sources, timestamps, owners, propagation targets and failure alerts.
Where the data comes from
For each field, record source, collection notice, purpose, definition, allowed values, timestamp, refresh, transformations, recipients, retention, deletion, correction and owner. Trace the path from source system through identity matching to the segment.
Label data types. Declared data comes from the person, such as a topic choice. Observed data records an action. Derived data is calculated. Inferred data predicts, and must never be presented as fact.
One address can mean a household, a role, an alias or a recycled account. Uncertain stitching should not carry high-impact targeting. Consent records and confirmed addresses are the raw material of email list building, so trace their lineage first.
Verify before release
- Run positive, negative, boundary, missing-data, stale-data, overlap and suppression tests on sampled records with an authorized reviewer.
- Reconcile counts against source totals and the previous version, then dry-run to a non-sending destination.
- Segregation of dutiesthe GAO standards for internal control are the reference for separating rule authorship from approval and retaining evidence of both.
- Purpose and minimizationthe W3C privacy principles support asking whether each field is genuinely needed for the stated decision.
- Platform fitthe GOV.UK guidance on choosing technology frames the test as your own requirements, covering suppression priority, time zones, versioning and exit.
Each check ends with a recorded result and a named approver.
Fairness and group size
Compare who is included, excluded, misclassified, over-contacted or denied useful information. Ordinary fields act as proxies: geography, purchase history, names and model scores can reproduce distinctions nobody intended.
Keep the review to one question set. Purpose, necessity, accuracy, benefit, harm and safer alternatives. Then add a challenge route and a correction path for affected people.
Tiny groups need protection as well as review. Set minimum sizes, restrict row-level access, aggregate reports, and never pad a send to meet a quota.
Measure the decision, not the click
Ask whether the segmentation decision improved an outcome against a defensible alternative. Randomize eligible units where practical, keep a holdout, define exposure, set the analysis window, and predefine the primary outcome and guardrails.
Track delivered eligible contacts, task completion, renewals, retained value, unsubscribes, complaints, support contacts and returns. Opens are noisy. Record uncertainty, and do not read association as causation.
Where the rule runs inside a platform, email automation covers the triggers, delays and frequency machinery that can override a segment without warning.
