Measurement
Measuring one customer data journey without pretending correlation is cause
Measure one England customer-data journey with precise event definitions, visible identity uncertainty, correction states and a clear boundary around causality.
Measure a customer data platform by the decision it supports, not by its profile store size.
For an England buyer, the useful unit is one controlled journey whose events, identity rules, exclusions and correction states are known.
This guide uses a fictional membership organisation. It receives authorised account events, reviews possible identity matches, preserves objections and suppression, and sends an approved audience to a destination.
There are no production records or results in this article. Every threshold and dashboard cell remains blank until the buyer supplies accepted evidence. A named statistician, evaluation specialist, data governance lead, UK privacy and PECR lawyer, security reviewer, accessibility specialist and finance reviewer must approve the relevant definitions before use.
What to take away
- Measure a customer data platform by the decision it supports, not by profile store size.
- Keep raw events, identity decisions, corrections, dashboard numbers, attribution credit, financial records and causal estimates separate.
- Write a measure card with decision, owner, event, population, numerator, denominator and clock before calculating.
- Treat preferences and permissions as recorded states, not as proof that every intended use is permitted.
- Never sum attribution credits from several platforms as though they were mutually exclusive.
Keep the systems and evidence layers apart
A CDP label says little about what generated a figure. The data journey may touch a CRM, warehouse or lake, identity-resolution service, consent or preference system, tag or device storage and access technology, analytics service, activation platform, clean room, managed integration and final destination. Record each as a separate node with its legal entity, function, schema and version.
Seven Evidence Layers
Raw & identity
- Raw event
- Named source
- Identity decision
- Rule and uncertainty
- Correction state
- Objection or deletion
- Dashboard number
- Documented definition
- Attributed credit
- Declared rule
- Reconciled finance
- Finance owner
- Causal estimate
- Counterfactual
Corrections & dashboards
- Raw event
- Not a count
- Identity decision
- Not a customer
- Correction state
- Not a profile
- Dashboard number
- Not reconciled
- Attributed credit
- Not revenue
- Reconciled finance
- Accepted record
- Causal estimate
- Not correlation
Credit & causal
- Raw event
- Not causal
- Identity decision
- Not causal
- Correction state
- Not causal
- Dashboard number
- Not causal
- Attributed credit
- Not causal
- Reconciled finance
- Not causal
- Causal estimate
- Supported only
The report should then separate seven kinds of evidence:
- A raw observed event from a named source.
- An identity decision, including the uncertainty and rule used.
- A correction, objection, suppression or deletion state.
- A number shown in a supplier dashboard under its documented definition.
- Credit allocated by a declared attribution rule or model.
- A reconciled financial record accepted by the finance owner.
- A causal estimate supported by an appropriate counterfactual.
Do not convert one layer into another. A profile count is not a customer count. A destination-reported conversion is not reconciled revenue. Attributed credit does not establish what would have happened without the activity.
HM Treasury's Magenta Book, updated on 15 May 2026, distinguishes monitoring and process evidence from impact evaluation, which asks whether outcomes can be attributed to an intervention. It is central-government guidance, not a private CDP standard, but the distinction is directly useful here.
Start with a decision and write a measure card
The owner should be able to complete this sentence: "If the measure crosses [buyer-approved threshold] during [period], I will [response], unless [exclusion]." A dashboard without a decision rule invites commentary rather than action.
Every measure card needs the following fields:
| Field | Required entry |
|---|
Measure Card Fields
- Decision and owner
- Exact event and eligible population
- Numerator, denominator and unit
- Clock, date range and Europe/London timezone
- Source, query, schema and model version
- Identity and deduplication rule
- Late, missing, unmatched and duplicate treatment
- Exclusions and correction status
Keep a frozen copy of the definition beside the result. If a supplier changes its event meaning or the buyer changes an identity rule, the series has a break. Recalculate only when the owner can reproduce the earlier query and explain what changed.
The Office for National Statistics describes statistical quality as fitness for purpose across dimensions including relevance, accuracy, timeliness, accessibility and comparability in its quality definition. That page concerns official statistics. It does not certify a commercial dashboard, but it offers a disciplined set of questions for the buyer's evidence record.
Define events before calculating rates
An event should state what happened, where it was recorded and which timestamp governs inclusion. "Audience sent" might mean a package created, accepted by a destination or made available for use. Those are three different events. Choose one.
Three Meanings of Audience Sent
Package created
- Event recorded
- Yes
- Destination confirmed
- No
- Usable by destination
- No
Accepted by destination
- Event recorded
- Yes
- Destination confirmed
- Yes
- Usable by destination
- No
Made available for use
- Event recorded
- Yes
- Destination confirmed
- Yes
- Usable by destination
- Yes
For a rate, the numerator must be a subset of the denominator. The eligible population should say whether corrected records, unresolved identities, refused tracking, suppression and late events are included. Use Europe/London for the reporting period and keep source timestamps separately so daylight-saving and delayed ingestion can be audited.
Identity needs particular care, so report proposed joins, accepted joins, rejected joins and unresolved candidates separately.
A higher match rate can be made to look attractive simply by forcing uncertain identities together. The ICO's accuracy guidance says keep source and status clear and consider challenges to accuracy. The page is under review after the Data (Use and Access) Act 2025; recheck before publication.
Treat preferences and permissions as evidence, not shorthand
A preference field is a recorded state from a named source at a particular time. It is not proof that every intended use is permitted. Preserve purpose, channel, collection route, version and later objection or suppression. Do not overwrite a refusal with a newer import merely because the newer record has a later timestamp.
If measurement uses pixels, links, device identifiers or similar technologies, the ICO's final storage and access technologies guidance dated 29 April 2026 needs a separate review. Collecting fewer fields can also improve clarity. The ICO's data minimisation guidance frames the question as whether personal information is adequate, relevant and limited to what is necessary for the stated purpose.
These are UK-wide regulator sources, including England. They do not determine a lawful basis, PECR exception or retention period for the fictional journey.
Build the dashboard around exceptions
The first view should tell the decision owner whether the evidence is usable. Show source freshness, schema and query versions, failed ingestion, unresolved identity candidates, correction backlog, unpropagated suppression, rejected destination rows and withdrawn figures. Put the decision metric below that evidence-health view.
The Government Analysis Function's dashboard guidance is aimed at government analysts. It treats accessibility, security and the ability to maintain a dashboard as selection concerns. A buyer can apply the same practical questions without suggesting that its internal CDP report is an official statistic.
Reports need a correction route. When a source error or identity defect is found, freeze the affected output, record the old value and reason, rerun the accepted query against the corrected state, and publish the correction marker. Do not silently replace a number. The owner also needs a withdrawal route when the result can no longer support the decision.
Choose an attribution statement that matches the method
Manual observation may say that an event occurred after a declared source. Deterministic rules can allocate credit when identifiers and windows meet written conditions. Supplier models may allocate credit using methods documented by that supplier. Matched observational analysis can adjust for measured differences, but unmeasured confounding may remain. Statistical association describes a relationship; it is not a causal conclusion.
An experiment or credible quasi-experimental design may estimate incremental effect if the counterfactual, assignment or identifying assumptions, interference, attrition and uncertainty are defensible. The Magenta Book discusses experimental, quasi-experimental and theory-based methods. A qualified evaluator should choose and review the design for the actual decision.
Never sum credits from several platforms as though they were mutually exclusive. Each may claim the same outcome under a different identity graph, window or model. Reconcile the underlying event once, retain each allocation separately and label it supplier-reported or rule-based.
Protect the reporting chain
Queries, exports and correction logs may contain personal or commercially sensitive information. Restrict access, separate service identities from human accounts and keep audit information sufficient to investigate a change. The NCSC's guidance on using SaaS securely covers activity logs, incident response and recovery for UK organisations. It does not validate a particular supplier or report.
Version the transformation code or documented manual procedure. The Government Analysis Function's reproducible analytical pipelines strategy advocates version control, testing, documentation and auditable processes for government analysis. A private buyer can adopt those techniques while recognising the source's public-sector scope.
Do not borrow a benchmark without admission tests
No authoritative England or UK benchmark for this defined CDP journey was identified in the sources reviewed on 6 September 2026. Broad software, advertising or business statistics cannot fill that gap. A supplier average may use a different product edition, population, identity rule, event definition or customer selection process.
Admit an external comparison only if it discloses the journey and function, geography, field dates, population and sampling frame, numerator and denominator, identity and deduplication rules, model version, exclusions, weighting, uncertainty, funding and missing-data treatment. Even then, show it beside the buyer's measure rather than substituting it for a threshold.
The UK Statistics Authority's Code of Practice for Statistics, edition 3.0 applies to official statistics and promotes suitable sources, rigorous methods and openness about quality. It is not authority for calling a commercial metric an official statistic. Its source-admission discipline is useful precisely because most niche benchmark claims lack comparable definitions.
The publication gate
To leave HOLD, the buyer should supply a versioned data dictionary, complete measure cards, approved privacy and PECR analysis, query and schema records, and identity and correction logic. They should provide finance reconciliation, accessibility findings, and an attribution statement proportionate to the method.
The statistical reviewer checks uncertainty and comparability. The finance owner confirms each amount belongs to the stated entity and period.
The next action is to choose one decision, complete one blank measure card and reproduce it from authorised fixtures. If the team cannot explain every state between source event and reported value, the measure is not ready for operational use.
Before you act
- Define the decision and owner before choosing any metric.
- Write a measure card with exact event, population, numerator, denominator and clock.
- Separate raw events, identity decisions, corrections, dashboard numbers and causal estimates.
- Preserve objections and suppression; do not overwrite a refusal with a newer import.
- Freeze the definition beside the result and record any change.
- Reconcile the underlying event once before summing attribution credits.
Common questions
What is the useful unit for measuring a customer data platform for an England buyer?
The useful unit is one controlled journey whose events, identity rules, exclusions and correction states are known. The article says to measure a CDP by the decision it supports, not by the size of its profile store. Every threshold and dashboard cell remains blank until the buyer supplies accepted evidence.
Why should a profile count not be treated as a customer count?
A profile count is not a customer count because the report separates seven kinds of evidence, including raw observed events, identity decisions and dashboard numbers. Do not convert one layer into another. A destination-reported conversion is not reconciled revenue, and attributed credit does not establish what would have happened without the activity.
How should identity matches be reported?
Report proposed joins, accepted joins, rejected joins and unresolved candidates independently. A higher match rate can be made to look attractive simply by forcing uncertain identities together. The ICO accuracy guidance says organisations should keep source and status clear and consider challenges to accuracy.
In this guide
- Customer data platforms key metrics: defining one objection metricCustomer data platforms key metrics for one England journey: a renewal objection metric defined field by field, edge cases counted, threshold left blank.
- What a customer data platform reporting dashboard should show about evidence health and correctionsA blank customer data platform dashboard that leads with evidence health, covering freshness, identity exceptions, corrections, definitions and decision rows.
- Attribution methods for customer data platforms and when to withhold causal claimsAttribution methods for customer data platforms, covering declared, deterministic, modelled, matched and experimental routes and when to withhold causal claims.
- Three customer data platforms measurement mistakes that overstate a resultCustomer data platforms measurement mistakes that overstate a result: a rising match rate, supplier credit called cause and totals counted twice.
- Screening customer data platform benchmark claims before you trust a numberScreen CDP benchmark claims through a reproducible protocol for population, identity rules, denominators, uncertainty and funding, with no invented England result.
