Careers
Email marketing analytics: what to keep and what to drop
Email marketing analytics in 2027 connects delivery, interaction, conversion, cost, privacy, experiments, and mature value through definitions and evidence.
What to take away
- Keep four thingsacceptance counts with a stated denominator, negative signals, task completions reconciled to finance, and randomized incrementality results.
- Drop open rate as an engagement score, click-to-open rate as standalone proof of health, and platform-attributed revenue presented as incremental.
- Write the decision, eligible population, alternative, outcome, guardrails and data cutoff before you shop for a tool.
- Version formulas, identifiers and maturity windows so a chart from March still means the same thing in December.
Set up the analytics around the decision, then keep only the metrics that carry a defensible denominator and a named owner.
The decision contract comes first
Write the decision down: send, stop, expand, reduce, redesign, retime, change audience, investigate, or retire. Record the owner, the eligible population, the alternative, the primary outcome, guardrails, the horizon, the data cutoff, and the action for each result.
Decision contract fields
- Decisionsend, stop, expand, reduce, redesign
- Decision owner named
- Eligible population defined
- Alternatives listed
- Primary outcome and guardrails
- Time horizon and minimum effect
- Data cutoff and action per result
Notes on email segmentation from the field shows how the eligible population rule changes every number that follows it.
The GOV.UK Service Manual's guidance on choosing digital analytics tools lists need, cost, storage, access, data quality, exports, privacy and transaction measurement. One concrete check follows: can this tool supply the evidence the decision needs?
The keep list and the drop list
| Metric | Keep or drop | Reason |
|---|---|---|
| Accepted messages | Keep | Provider acceptance with a stated denominator |
| Hard bounces and policy blocks | Keep | Drive suppression and diagnosis |
| Complaints by mailbox provider | Keep | Gmail Postmaster defines its own rate |
| Task completions reconciled to finance | Keep | Matches accounting scope and period |
| Open rate as an engagement score | Drop | Apple prefetches remote content |
| Click-to-open rate standing alone | Drop | Inherits every open problem |
| Platform-attributed revenue as incremental | Drop | A model output, not a counterfactual |
| Blended engagement score | Drop | Hides the parts inside it |
Pick the unit of analysis before you average anything. A row may be a send attempt, a delivered message, a contact, an account, or a day. Never divide message-level clicks by person-level subscribers without a stated transformation.
Example: a delivery denominator you can defend
Take an invented illustration, not a benchmark. A campaign logs 10,000 send attempts: 9,700 accepted, 180 deferred, 120 rejected. The honest guide to email deliverability covers the placement questions that acceptance cannot answer.
A delivery denominator you can defend
- Set the denominator as send attempts recorded by the sending platform.
- Set the numerator as messages the provider accepted with no rejection code.
- Split the remainder into permanent bounces, temporary deferrals, policy blocks, authentication failures and unknown states.
- Refresh the count, because a deferral can become a delivery or a bounce hours later.
- Keep the raw log so another analyst can rebuild the figure.
For this invented case the delivery rate is 9,700 divided by 10,000, or 97 percent. Delivery means provider acceptance, not inbox placement or reading.
Delivery states to separate
Counts as delivered
- Accepted by provider
- Yes
- Permanent bounce
- No
- Temporary deferral
- No
- Policy block
- No
- Authentication failure
- No
- Unknown state
- No
Does not count as delivered
- Accepted by provider
- No
- Permanent bounce
- Yes
- Temporary deferral
- Yes
- Policy block
- Yes
- Authentication failure
- Yes
- Unknown state
- Yes
Opens and clicks as diagnostics
An open records a remote-content request. Apple states that Mail Privacy Protection downloads remote content whether or not the person reads the message, so open counts can climb while attention stays flat. Do not suppress subscribers on opens alone.
Clicks arrive dirty. Security scanners, preview generators, forwarded recipients, bots and accidental taps all register. Keep raw events, then build one documented filtered view using timing, user agent, network and known scanner patterns. What matters most in email automation covers which platform creates which event before you filter it.
Name the formula. Click-through rate may divide unique clicking recipients by accepted messages, or by send attempts. Click-to-open rate inherits every open problem. Show numerator and denominator beside each percentage.
Attribution versus incrementality
Last click, first click, linear, position-based, time-decay and platform-defined models answer different questions. Consent, cross-device behavior and offline activity change the results. Report attributed revenue under a named model, and never present it as incremental impact.
Randomize where you can. Assign eligible accounts or households to the message decision and to a responsible alternative, preserve the assignment, and prevent crossover.
A holdout estimates what happens without the tested decision, but withholding may be improper for transactional or legally required mail. Define eligibility and duration, and set the maturity window before you look.
Identifiers, joins and consent records
Assign stable identifiers to campaigns, messages, links, audiences, experiments and journeys, and keep provider identifiers as mappings rather than keys. Govern naming, because Google Analytics campaign parameters are case sensitive.
Consent and opt-out records belong in the same dataset as performance data. Keep unsubscribe requests, completed preference changes and provider complaints as separate fields. Canadian consent rules and US opt-out rules attach to the message, not to the dashboard.
Join carelessly and revenue inflates. Shared addresses, aliases, guest checkouts and account merges duplicate or misassign outcomes. Document join keys and unmatched rates.
Checks before a number goes live
- Test schema, required fields, uniqueness, event order, timestamp drift, duplicates, late arrivals and join rates on every refresh.
- Reconcile revenue to finance for the same scope, period and currency rules.
- Quarantine anomalies instead of deleting them, and keep raw events for review.
Privacy sits inside that control set. The W3C's Privacy Principles call for purpose limitation and data minimization. Applied here: no personal data in URLs or analytics keys, and row-level access limited to named roles.
The GAO's Standards for Internal Control in the Federal Government treat documentation and monitoring as control activities. Applied here: every published metric has a written definition and an owner who can reproduce it.
A measurement plan cannot repair an undocumented decision. Getting email marketing strategy right the first time explains why a test only answers the question the strategy wrote down.
