Marketing attribution key metrics: a documentation framework

Marketing attribution key metrics should describe a decision and a counting rule, not a fashionable dashboard label. Define the event, eligible population and evidence boundary first. Leave the value blank until the required records pass quality and authority checks.

This method was researched on 6 September 2026 for attribution operated in England. It provides no England benchmark and no legal or statistical conclusion. UK privacy and PECR questions need qualified review against the real collection design.

What to take away

Define the event, eligible population and evidence boundary before filling in any attribution metric card.
Operational health measures show whether collection worked, but they do not show marketing effectiveness.
Keep modelled estimates and causal experiments separate, each with its own assumptions and uncertainty.
Give every metric card an expiry event and revisit it when schemas, triggers, keys or rules change.
Publish the card beside the result so users can see population, period, source and causal limits.

Structure the metric card

Use one card per measure:

Common marketing outcome metrics

Define the business meaning and scope of each measure before reporting it:

ROAS (return on ad spend):revenue credited to advertising divided by ad spend under the stated attribution rule and period.
ROI (return on investment):return relative to cost; specify whether the return means attributed revenue, profit or incremental value.
CPA (cost per acquisition):campaign cost divided by the defined acquisition events.
CAC (customer acquisition cost):the included acquisition costs divided by the new customers acquired; state the cohort and costs included.
Conversion rate:conversions divided by the eligible denominator, such as sessions or clicks; name that denominator.
Revenue per session:revenue divided by eligible sessions.
LTV (lifetime value):observed or estimated customer value over a stated horizon and scope.
Assisted conversions:conversions for which a channel is recorded as an assist under the platform's path rules; this is not, by itself, evidence of incremental impact.

For example, a ROAS card can specify credited revenue as the numerator and ad spend as the denominator, then record the eligible campaign population, period, platform, credit rule and extraction date. Enter a value only when the underlying records pass the stated checks.

Metric card required fields

Decision and owner
Event and trigger
Eligible population
Numerator and denominator
Geography and timezone
Source and model version
Uncertainty and expiry

Field / Required entry

Decision and owner: ____
Event and trigger: ____
Eligible population: ____
Numerator: ____
Denominator or unit: ____
Geography rule: ____
Period and Europe/London timezone: ____
Source and extract version: ____
Query, configuration or model version: ____
Late, duplicate and missing treatment: ____
Unmatched, refunded and corrected treatment: ____
Uncertainty and expiry: ____

Reject a rate if the denominator is missing or changes between periods. A count should say whether the entity is a person, device, session, lead, order or another defined unit.

Separate operational health from attribution

Operational measures show whether collection and processing worked. Useful blank forms include:

Operational health vs attribution

Operational health

Question: Did collection work?
Example: Accepted event share
Shows effectiveness: No
Key output: Rejected, late, unavailable

Attribution

Question: What value is credited?
Example: Platform credit allocation
Shows effectiveness: No, only assigned value
Key output: Rule, lookback, source period

accepted event share = accepted eligible events / received eligible events

unmatched record share = unmatched eligible records / records presented for matching

correction share = corrected released records / released records eligible for correction

These measures do not show marketing effectiveness. State which records were rejected, late or unavailable. The Government Data Quality Framework is for the public sector, but its distinction between completeness, uniqueness, consistency, timeliness, validity and accuracy helps prevent one vague quality score.

Label descriptive and credited outcomes

General analytics can report a defined event within known coverage. Cross-system reconciliation can report agreement and exceptions. Platform credit allocation can show the value assigned under a named rule and configuration. None of those results establishes why the event occurred.

Attribution models to name

State which model assigns credit, since the same customer path can produce different results under different rules:

First-touchassigns credit to the first eligible touchpoint; last-touch assigns it to the last eligible touchpoint.
Lineardistributes credit across eligible touchpoints; time-decay gives more credit to more recent touchpoints.
Position-basedgives different credit to selected positions in the path; document the actual weighting rule.
Data-drivenuses a model to allocate credit; record the platform's model version and available assumptions or limitations.

Lookback windows, eligible channels and implementation details can vary. Record the actual configuration rather than treating model names as interchangeable.

Preserve per credit line

Platform
Account scope
Rule
Lookback settings
Source period
Extraction date

For each platform-credit line, preserve the platform, account scope, rule, lookback settings, source period and extraction date. Do not combine credits from different platforms unless a documented reconciliation prevents double counting and preserves incompatible definitions.

Examples of platforms teams may need to document include Google Analytics 4 (GA4), Adobe Analytics, HubSpot and Salesforce. Do not assume their labels or configurations are equivalent: record the actual account, reporting source, conversion event, credit rule, lookback setting and extraction date for each line.

Where collection uses tags, pixels or similar technologies, the ICO's final storage and access technologies guidance requires a purpose-specific assessment. The guidance does not choose the right route for a particular setup. Report the observable population rather than pretending it covers everyone eligible for the business decision.

Keep modelled and causal measures distinct

MMM reports an estimate from a declared specification, input period and model version. Its card should include assumptions, diagnostics, uncertainty and the population to which interpretation is limited.

For marketing mix modelling (MMM), document the outputs the specification produces. These commonly include channel contribution, baseline, adstock (the carry-over effect of past media) and saturation (diminishing returns as media exposure or spend increases); definitions and estimates depend on the model.

An experiment or quasi-experiment needs the intervention, comparison, outcome, assignment or identification method, analysis plan, implementation evidence and uncertainty. HM Treasury's QPIE guidance explains in government evaluation that monitoring change is not enough to identify cause and that a credible counterfactual matters.

Expire and correct the definition

Give the card an expiry event rather than assuming its meaning will persist. Revisit it when a source schema, event trigger, identity key, consent route, refund rule, query or model changes. Preserve the prior version and state whether the new result is comparable.

A correction record should identify the affected release, original and revised definitions, reason, approval time and decision users notified. If the team cannot reproduce the earlier calculation, withdraw the figure instead of silently replacing it. Keep the correction visible wherever the original value informed action.

Attach a decision rule

State the buyer's proposed action, threshold, evidence owner and expiry. Keep the threshold blank until approved. A decision rule should say what happens when data are incomplete, uncertainty crosses the accepted boundary or the method version changes.

Publish the metric card beside the result. If the user cannot see population, period, source, exclusions and causal limit, the metric is not ready for decision use.

Before you act

Define the event, eligible population and evidence boundary first.
Reject any rate whose denominator is missing or changes.
State which records were rejected, late or unavailable.
Preserve platform, rule, lookback and extraction date for each credit line.
Record the affected release, revised definitions, reason and approval time.
Keep the threshold blank until the buyer approves it.

Common questions

What should a metric card contain before a value is entered?

One card per measure should record the decision and owner, event and trigger, eligible population, numerator and denominator. It should also record geography rule, period, source version, query version and uncertainty. Leave the value blank until records pass quality and authority checks.

Why should operational health measures be kept apart from attribution?

Operational measures such as accepted event share, unmatched record share and correction share show whether collection and processing worked. They do not show marketing effectiveness. Separating them stops a vague quality score from being read as evidence that a campaign caused an outcome, and it forces teams to state which records were rejected, late or unavailable.

When should a metric definition be revisited or withdrawn?

Revisit the card when a source schema, event trigger, identity key, consent route, refund rule, query or model changes. Preserve the prior version and state whether the new result is comparable. If the team cannot reproduce the earlier calculation, withdraw the figure instead of silently replacing it, and keep the correction visible wherever the original value informed action.