Four ways to attribute multilingual social results and what each can honestly claim

Multilingual social marketing attribution methods answer different questions. Monitoring describes observed events. A rule allocates credit. Contribution analysis tests a causal story. Experimental or suitable quasi-experimental evaluation estimates impact against a counterfactual. There is no universal winner.

What to take away

Common comparison frame

Compare one England campaign decision, eligible population, intervention, outcome event and period. Use the same asset versions, language strata, cost boundary and verified business outcome. For each method, record data sources, identity and joining rules, exclusions, missingness, timing, assumptions, uncertainty, owner and decision rule. Do not infer an individual's language or culture from profile attributes.

The 2026 Magenta Book is central-government guidance rather than a marketing standard, but it provides clear distinctions between monitoring and impact evaluation.

Descriptive monitoring

Question: what events were recorded? Inputs: versioned operational, delivery, response or conversion logs. Unit: a defined count, duration or rate with an eligible denominator. Strength: quick diagnosis of delivery and data quality. Failure mode: changes may reflect seasonality, audience mix, other activity or tracking changes. It cannot establish that multilingual work caused an outcome.

Use monitoring when the decision concerns release fidelity, missing data or whether a process is operating. Keep platform-reported values under the supplier definition.

Rule-based attribution

Question: which recorded touchpoint receives credit under the chosen rule? Inputs include touchpoints, conversion event, identity match, attribution window and tie-breaking logic. The output is attributed events or value, not incremental effect. Results change when the window, eligible channels or identity resolution changes.

This method can support finance reconciliation and operational allocation if the rule is stable and disclosed. It fails when unmatched records, cross-device activity or incompatible platform definitions are hidden.

Contribution analysis

Question: is the proposed causal pathway consistent with multiple sources of evidence? Inputs include a theory of change, process evidence, observed outcomes, rival explanations and stakeholder or research records collected with suitable permissions. The result is a qualified contribution claim, not automatically an effect size.

Use it when a controlled counterfactual is infeasible but the buyer needs to assess assumptions. Failure follows if analysts seek only confirming evidence or treat participant accounts as representative without a sampling basis.

Counterfactual impact evaluation

Question: what difference did the intervention make compared with what would otherwise have happened? Experimental designs use random allocation where feasible; quasi-experimental designs seek a credible comparison under explicit assumptions. Inputs must include comparable outcome definitions, allocation or selection process, contamination checks, power or precision work, attrition and pre-specified analysis.

The current QPIE guidance says outcome monitoring alone cannot identify causation and links stronger confidence to a strong comparison or control. Even a strong design may not generalise from one language, market or period.

Freeze the evidence plan

Before observing outcomes, write the objective, eligible population, intervention and comparison, primary event, numerator and denominator or unit, analysis period, timezone, data sources, query versions, exclusions, missing-data treatment and uncertainty method. Record language and market strata from approved assets or self-described research data, never profile inference. Name the analyst, decision owner and the result that would weaken the proposed explanation.

Reconcile operational delivery before interpreting response. If the target asset was not approved, the comparison received different treatment or tracking failed unevenly, the prerequisites have failed. Report that failure instead of choosing another attribution model after the event.

Selection rule

Choose the least burdensome method that answers the decision honestly. Use monitoring for operations, a declared rule for allocation, contribution analysis for a tested causal account, and counterfactual work for an effect question where design conditions can be met.

Privacy, PECR, research ethics and accessibility remain separate. Translation, advertising, finance review stay separate; if evidence cannot answer causality, report association and stop.

Before you act

Common questions

What is the difference between monitoring and impact evaluation?

Monitoring describes observed events, such as versioned operational or conversion logs, and can diagnose delivery and data quality. It cannot establish that multilingual work caused an outcome. Impact evaluation instead estimates a difference against a counterfactual, which the Magenta Book distinguishes from monitoring.

When should I use contribution analysis instead of a controlled experiment?

Use contribution analysis when a controlled counterfactual is infeasible but the buyer still needs to assess assumptions. It tests whether the proposed causal pathway is consistent with multiple sources of evidence, including rival explanations. The result is a qualified contribution claim, not automatically an effect size.

What must be written down before observing outcomes?

Write the objective, eligible population, intervention and comparison, plus the primary event and unit. Also record the analysis period, timezone, data sources and query versions. Add exclusions, missing-data treatment and uncertainty method. Record language and market strata from approved assets or self-described research data, never profile inference. Name the analyst, decision owner and the result that would weaken the explanation.