Rules
Part of Customer analytics: what to keep and what to drop
Customer analytics questions nobody wants to ask out loud, and straight answers
Candid customer analytics questions on identity, permission, coverage, causation and data quality, each with a direct answer, followed by the working definitions and checks that keep the analysis reviewable.
Six steps make up the review routine behind the answers below.
- Choose an approved customer unit and name the decision the analysis should change.
- Trace one record from its source through profile and segment to a possible action.
- Define the labels, eligibility rules, and periods used to read the results.
- Recheck permission, data quality, coverage, and corrections when the analysis is used.
- Mark whether the evidence describes behavior, predicts an outcome, or supports a causal claim.
- Record an owner, limits, review date, and reasons to revisit the decision.
The uncomfortable questions, answered
Is this record really one person?
Did these customers agree to this use?
Who is missing from the data?
Would they have bought anyway?
Is the data clean enough to act on?
What changes if the number turns out wrong?
Anchor the unit first
Choose one approved unit, such as a person, household, or account, and state the decision the analysis should change. This boundary keeps a profile from being mistaken for a complete account of a customer. Record the strategy choices before selecting a method or tool.
Write the intended business action beside that decision, then note who may receive it. If a team cannot name what could change after seeing the result, pause the work and clarify the request. This prevents a metric or platform from silently choosing the question on the team's behalf.
Follow one record
Follow a single record from its source through profile and segment to a possible action. That route reveals which identity rules shaped the record, whether permission was checked, and where data moved. If its inclusion cannot be explained, hold activation rather than treating the profile as self-explanatory.
Source to Profile to Action
- Source record with included and missing sources
- Identity state: anonymous, person, household, account
- Match confidence and unresolved conflicts
- Consent, purpose, suppression, deletion dates
- Profile freshness and outcome maturity
- Segment qualification and activation decision
- Correction route and restatement of affected audiences
Use a reviewer-facing example that shows the path without exposing unnecessary personal information. Masked values can still demonstrate why a record qualified. Keep attention on the transition from source to action, and avoid presenting the derived profile as complete or certain.
Illustrative example, with invented values. Record 48213 enters from the web-order source as a logged-in purchase on 3 March. A hashed email matches a loyalty account, so it joins profile P-77. Rule R4, "purchased in the last 90 days", places it in the segment "recent buyers". The planned action is a replenishment email.
At each step the reviewer asks one question. Did the identity rule allow the match? Does the consent flag cover marketing email, not only order updates? Is the order inside the 90-day window? Here all three answers are yes, so the record qualifies. Had the flag covered order updates only, the record would drop out and the audience would shrink by one.
Make labels testable
Set a unit, rule, and period for every label that affects eligibility or interpretation. Stable definitions tell readers who is included and what a result can mean.
They also make it easier to spot when terms have shifted between analysis and the decision, rather than assuming a label carries one fixed meaning.
| Question | Working definition |
|---|---|
| What is a customer? | The approved person, household, or account unit |
| What is a profile? | A derived view governed by documented identity rules |
| What is a cohort? | Customers sharing a defined entry condition and period |
| What is a segment? | A group selected by a rule or model for a purpose |
| What is retention? | Eligible customers active under a stated event and window |
| What is lifetime value? | Realized or forecast economics over a defined horizon |
| Can analytics establish cause? | Only with a credible counterfactual design |
| Who owns action? | A named business owner, with data and method stewards |
Treat these working definitions as a data-quality check, not a requirement that every organization adopt identical vocabulary. Keep identity rules beside the source data so a reviewer can see how a profile formed. A device ID identifies a device or browser unless a documented process connects it to another approved unit.
A customer may qualify for more than one segment unless the rules make groups exclusive. Measure segment overlap before treatment, and state whether overlap is allowed. Otherwise, readers may assume separate categories even though the same customer can appear in several groups.
Repeat the eligibility checks
Repeat permission, data quality, and coverage checks when the analysis is used, not only when it begins. Customer status, identity, preferences, returns, or service outcomes may change; therefore, an audience that qualified earlier may need correction before a later decision or action.
Name who can raise a correction and how the decision will be reconsidered. Keep the correction route and effective date beside the answer. When a number is restated, send the updated result to everyone relying on it, so an earlier result does not remain in circulation as though nothing changed.
Review data flow against applicable permission requirements, contract terms, and the configured process; a reference cannot replace those checks. NIST's privacy guidance is one official reference, while the record should still state the approved purpose and the limits of the review. Include the permission review in that record.
Which rules apply depends on where the customer lives and what the audience is used for. In Canada, PIPEDA requires meaningful consent for each purpose, and Canada's anti-spam law (CASL) requires consent before commercial electronic messages. Quebec's Law 25 adds its own consent requirements for residents of that province.
In the United States, the CAN-SPAM Act requires honoring an email opt-out within 10 business days. California residents can opt out of the sale or sharing of their personal information, and businesses must treat a Global Privacy Control signal as that opt-out. Counsel should confirm which rules reach your audience.
Set an evidence ceiling
Classify the result as a description of observed behavior, a prediction, or a causal claim. State the eligible population, evidence period, and method alongside the claim. That boundary keeps an observed pattern or forecast from being presented as proof that a treatment caused an outcome.
Predictions can guide review, but a causal conclusion requires a credible counterfactual design, not merely a difference between groups. The analysis checklist connects the proposed method with its data and decision requirements. State plainly what the evidence cannot establish, particularly when readers may use it to authorize an action.
Leave a reviewable trail
Keep a record that another reviewer can test: the unit, eligible population, evidence period, method, intended action, and limits. Include the definitions and the person accountable for the decision. A record should show why the result supported an action, not just preserve the final metric without its reasoning.
Name the business owner and the data and method stewards, then set a review date and the change that would prompt an earlier check. A written ownership record can help keep responsibility visible. The record should distinguish accountability from technical review.
Do not activate every profile automatically. Activation needs a purpose, permission, quality threshold, appropriate treatment, suppression rules, and an owner. The decision owner should be able to follow the reasoning without relying on memory or assuming every customer record is complete. Related questions are answered in marketing analytics strategy questions.
Set the activation thresholds before the first send, and adapt these example numbers to your own risk. A team might require a valid consent flag on every record, a duplicate rate under 2 percent, and no source older than 30 days. It then checks 50 records by hand and holds the audience if more than 2 fail.
When the same question returns, compare the underlying records before changing the metric. A changed population or source may explain a different result without showing altered behavior. Put the measurement contract beside the correction route and state what would trigger another review.
