Ninety days of attribution work, gated from framing to archive

Ninety days is an editorial planning convention, not a normal implementation time or performance promise. A buyer may need to pause, shorten or replace any phase. The useful outcome is a signed decision record: proceed, collect more evidence, change the question or stop.

Effort depends on source count, event complexity, review load and legal scope; no fixed cost applies because those factors vary.

What to take away

Days 1 to 15: frame one decision

Write the allocation or measurement question, owner, available options and review horizon. Define the outcome event, population, England geography rule, period, timezone and channel coverage. Preserve a no-change option.

Map how the proposed activity is expected to affect the outcome, then identify assumptions and outside influences. The Government Analysis Function's Theory of Change toolkit was designed for public interventions. Its focus on context, inputs, outputs, outcomes, causal links and risks is useful planning structure, not validation of a commercial campaign.

Gate: the decision owner accepts the question and the measurement lead confirms that each term can be recorded. Otherwise, return the brief.

Days 16 to 30: inventory evidence

List every source system, event field, owner, extraction method, query/version, transformation, join, missing value and correction process. Capture platform configurations and reporting windows. Do not merge channel credits.

Examples include Google Ads conversion actions, GA4 purchase and session_start events, and Shopify order webhooks; fields might be transaction_id, gclid, conversion_time and refund_ts. The actual list depends on the stack.

Classify the intended job as platform allocation, cross-system reconciliation, aggregate modelling, experiment or causal incrementality. Google's Ads attribution documentation shows that event time, network coverage, conversion source and lookback settings affect reports in that service. Store the actual settings rather than relying on the channel label.

Gate: provenance is reproducible and gaps are visible. If a business outcome cannot be tied to a stable definition, stop before model selection.

Days 31 to 45: clear independent risks

Ask separate reviewers to examine privacy and PECR, security, accessibility, advertising evidence, finance and supplier terms. If pixels, cookies or related device technologies are proposed, use the ICO's current storage and access technologies guidance as a source for the legal review. Do not assume a platform setting settles the organisation's purpose or duties.

Define access roles, test data, retention, deletion, incident route, export and rollback. Put unresolved issues in the decision log.

Gate: every non-compensating review records pass, conditional pass or stop, with evidence and expiry. A favourable measurement view cannot override another gate.

Days 46 to 60: design the evaluation

Choose the least complex method capable of answering the decision. Specify confounders, counterfactual, assignment or comparison rule, model/version, uncertainty analysis and transfer limits. HM Treasury's Magenta Book covers proportionate evaluation for government. This buyer plan adapts the discipline while retaining the source's public-sector scope.

Leave sample requirements, completeness limits and decision thresholds blank until the statistics reviewer and buyer approve them. Use synthetic fixtures to test transformations and access without fabricating an outcome.

Illustrative example only: treat results as directional after 30 conversions per arm and eligible for a proceed decision after 300 per arm; the right numbers depend on baseline rate, minimum detectable effect, traffic and clustering.

Gate: the pre-analysis record is signed before live results are inspected.

Days 61 to 75: run or pause

Freeze input versions, execute the approved process and log deviations. Keep observed outcomes, platform-attributed credit, model estimates and causal estimates in separate fields. If collection, assignment, source coverage or consent conditions depart materially from plan, pause rather than repair the story afterwards.

Gate: reviewers decide whether deviations permit analysis, require a revised plan or invalidate the run.

Days 76 to 90: decide and archive

Present the estimate with uncertainty, missing channels, negative findings and alternative explanations. ASA and CAP's substantiation guidance is relevant if the organisation wants to turn results into objective advertising claims; claim review remains distinct from the internal allocation decision.

Record the owner's decision, dissent, exposure, rollback route and next trigger. Archive the brief, approvals, data references, code, configurations, output and correction history. If evidence is inconclusive, say so. Day 90 does not turn an unresolved question into an answer.

Worked example record (illustrative):

Before you act

Common questions

What is the purpose of the ninety-day plan?

It sets a review rhythm and decision gates, not a fixed delivery date. A buyer can pause, shorten or replace any phase. The file ends with the signed record described above.

What must happen before live results are inspected?

Sign the pre-analysis record first. The statistics reviewer and buyer then fill the blank thresholds. Synthetic data can test the pipeline without creating a result.

What should be done if evidence is inconclusive?

Record the inconclusive finding as the outcome. The owner still signs a decision: continue, collect more evidence, change the question or stop. Archive the signed record and its supporting files.