Measurement
Part of Attribution measurement built for England around five separate jobs
Designing an attribution dashboard with a decision strip and an exceptions panel
Design an attribution dashboard around a real decision, with metric definitions, exception states, uncertainty, corrections and accessible evidence.
A marketing attribution reporting dashboard should help a named person make a declared decision. It should not simulate certainty with a row of unexplained totals. Put the evidence boundary, exception states and correction history on the main decision path.
Before using this template, supply a named decision owner, the job class, the evidence boundary, the population, period and unit, and qualified reviewers for the method and data. The values, alert levels and decision thresholds stay empty until those inputs exist. Research for this design took place on 6 September 2026; no interface, report, user or dataset was tested.
What to take away
- A dashboard should help a named person make a declared decision, not simulate certainty with unexplained totals.
- The default view should show the job class, owner, evidence expiry, population, period, unit, uncertainty and causal boundary before any metric total.
- Exception states belong in a separate panel, with unknown states left unknown and reconciliation shown as record agreement.
- Each visible result should link to its input manifest, code commit, configuration, model version and reviewer record.
- If evidence cannot be reproduced or a mandatory reviewer has not approved it, withdraw the view and preserve the prior release.
Start with the decision strip
The first screen or page should answer:
- What decision is this report allowed to inform?
- Who owns that decision and when does the evidence expire?
- Is the job platform credit, reconciliation, descriptive analytics, MMM, or experiment or quasi-experiment analysis?
- Which population, geography and period are covered?
- Is the current state draft, reviewed, released, corrected, withdrawn or HOLD?
Avoid a blended "marketing performance" heading. The five job classes produce different evidence and should not share one implied claim.
Place the metric card next to the value
For each result, show event, eligible population, numerator, denominator or unit, Europe/London timezone, source, extract, query, configuration and model versions. Add the rules for late, duplicate, missing, unmatched, refunded, reversed and corrected records.
Use a disclosure drawer only for extra detail. Population, period, unit, uncertainty and causal boundary belong in the default view. Government Analysis Function guidance on quality and uncertainty is written mainly for official-statistics producers, but its case for explaining limitations where numbers are used is relevant to internal reporting.
Give exceptions their own panel
Show eligible, received, accepted, rejected, duplicate, late, missing and unmatched records as separate states. Include the source-to-report reconciliation status and last successful correction. Leave a state unknown when the system cannot observe it.
The Government Data Quality Framework applies to public-sector data. Its separate dimensions for completeness, uniqueness, consistency, timeliness, validity and accuracy are a better prompt than one invented quality score.
A worked example with placeholder values appears below. The counts and thresholds are not benchmarks.
| Decision strip field | Placeholder value | Exception state | Placeholder count |
|---|---|---|---|
| Decision | Pause paid search when cost per accepted lead reaches the stop condition | Eligible | 10,000 |
| Owner | Growth lead named in the release record | Received | 9,500 |
| Job class | Reconciliation | Accepted | 8,700 |
| Evidence expiry | 30 days after release, or earlier if a source changes | Rejected | 300 |
| Population and period | England, 1 to 30 June 2026 | Duplicate | 200 |
| Unit | Accepted lead | Late | 150 |
| State | Draft, pending reviewer approval | Missing | 100 |
| Stop condition | TBD until the buyer names a threshold | Unmatched | 50 |
| Reconciliation status | 8,700 accepted matched to source; 800 unresolved |
Make the evidence reproducible
Link the visible result to its input manifest, code or query commit, configuration, model version, operator and reviewer record. Keep a plain-language method note for decision users. The Government Analysis Function's RAP strategy supports version control, input validation and documented manual steps in government analysis. These practices improve auditability but do not prove a commercial result.
Present attribution limits directly
Label platform-assigned credit with its rule and account scope. Display reconciliation as record agreement, not contribution. Describe MMM as a modelled estimate with assumptions. Use causal wording for experimental or quasi-experimental results only when the design and counterfactual support it.
A decision box should state the action, the buyer threshold, uncertainty, contrary evidence and stop condition. If the buyer threshold is blank, set the stop condition as a review date, a named person who can pause spend or reporting, and a test of whether contrary evidence overturns the current read.
If the causal question remains unanswered, say so rather than letting a chart imply the answer.
Test access and corrections
Do not rely on colour alone. Give charts text summaries, meaningful labels, keyboard access and a usable tabular export. GOV.UK's assistive-technology testing guidance is for government services. It provides useful task-based methods, but a private organisation must test its own configured dashboard and users.
Protect correction and access logs. The NCSC's cloud audit principle asks buyers to assess which operations are recorded and whether logs support investigation. It is not product assurance.
When a source, definition or method changes, preserve the prior release, mark the break in comparability and show the correction where the affected value appears. Withdraw the view if evidence cannot be reproduced or a mandatory reviewer has not approved it.
Before you act
- Name the decision and its owner before building.
- State the job class and evidence boundary.
- Show population, period, unit and uncertainty by default.
- List exception states separately, leaving unknown states unknown.
- Link results to manifests, code commits and reviewer records.
- Preserve the prior release when a method changes.
Common questions
What should a reader check before trusting a released view?
Check that the state is released, not draft or HOLD, and that a named reviewer has approved the method, data and thresholds. Check that the evidence is inside its expiry date and that the source-to-report reconciliation is complete or the unresolved states are visible. If any check fails, treat the view as an exception, not a result.
How should platform credit, reconciliation and MMM be labelled?
Label platform-assigned credit with its rule and account scope. Display reconciliation as record agreement, not contribution. Describe MMM as a modelled estimate with assumptions. Use causal wording for experimental or quasi-experimental results only when the design and counterfactual support it.
What should happen when a source, definition or method changes?
Preserve the prior release, mark the break in comparability and show the correction where the affected value appears. Withdraw the view if evidence cannot be reproduced or a mandatory reviewer has not approved it. The article also says to protect correction and access logs.
