Measuring multilingual social marketing without confusing attribution with cause

Multilingual social marketing measurement starts with a decision, not a dashboard. A useful report tells the reader what happened in a defined population and period, how the number was produced, what is missing and what decision follows. It does not turn delivery counts into audience comprehension or platform attribution into causal impact.

For England work, keep geography attached to every record. UK-wide regulatory guidance remains UK-wide; global platform data remains supplier data with its own definitions. Do not assign language, literacy, ethnicity, culture, identity or preference from names, locations, profiles or machine classifications. A target-language label describes an asset, not the person who saw it.

What to take away

Separate the layers before choosing metrics

Use a measurement chain with distinct evidence classes:

Movement down that list is not automatic. A completed translation does not show delivery. Delivery does not prove a person understood it. A click is not a purchase, and an attributed sale is not necessarily incremental.

The GOV.UK Service Manual guidance on setting performance metrics advises starting with objectives, identifying data sources, checking accuracy and giving measures context. It describes government digital services, so use its method rather than importing its service metrics into a campaign.

Write a metric card first

No number enters a report without a metric card. Record:

The AQuA Book, published in 2025, frames government analytical assurance around fit-for-purpose design, roles, verification, validation and sign-off. It is a useful discipline here, not proof that a marketing analysis is valid.

Choose a small decision-led set

An operations owner might need the share of eligible target assets with all required language, subject and accessibility approvals. The numerator is fully approved assets; the denominator is all assets eligible for those reviews in the stated period. Cancelled work must be treated explicitly. This measure says nothing about campaign response.

A publishing owner might track release fidelity: live items matching the approved copy, language, disclosure, destination and audience configuration divided by eligible releases. Capture the check time and platform identifier. A mismatch triggers pause or correction rather than a favourable average.

An audience-research owner may report a response distribution from participants who self-described the relevant language or format need. The sampling frame, invitation count, eligibility, non-response, consent, withdrawal and question version belong beside the result. Findings apply to the studied population, not everyone in England.

A commercial owner can reconcile defined conversions with finance records, using an agreed event and accounting period. Keep gross revenue, contribution and cash cost separate. The report should show unresolved matches and refunds rather than silently treating them as successes.

Build the data contract

For each input, name the legal and operational owner, system, field, definition, type, unit, timezone, permitted values, refresh schedule, cut-off, retention, access group and correction route. Add the language-asset identifier and source version, but do not attach inferred personal language labels.

Store raw extracts separately from transformed tables. Keep the extraction time, query, schema, transformation code or workbook version, and reconciliation result. When a source changes a field definition, preserve the old series and mark the break. Do not backfill silently.

The Office for Statistics Regulation's Code of Practice for Statistics requires official statistics producers to consider completeness, validity, accuracy, coherence, comparability and timeliness, and to explain limitations and uncertainty. Campaign reporting is not official statistics, but those quality questions expose weak evidence.

Make missingness visible

An empty value can mean no event, unavailable data, failed tracking, suppressed data, delayed reporting or an ineligible record. Give each state its own code. Publish the count and share of records in each missingness class. If missingness differs by language, channel or device, an overall average may mislead.

Do not fill gaps with zero unless the metric definition makes zero correct. Do not carry forward a prior value without labelling the imputation and testing its effect. The ONS explanation of survey uncertainty distinguishes sampling error from non-sampling problems such as non-response and processing error. Its statistical measures apply to suitable survey designs, not automatically to platform logs.

Design the dashboard as a reproducible view

Each tile should display the metric name, value, eligible denominator or unit, period, timezone, last refresh, source, version and quality flag. A reader should be able to open the definition and trace the displayed value to a frozen extract.

Show corrections and restatements. If late conversions arrive after the cut-off, publish the original and revised values with the rule that caused the change. A dashboard screenshot is not the audit record.

Use suppression and access controls where small cells or linked fields could expose people. The ICO's data protection by design guidance was updated in February 2026 after the DUAA and connects design with data minimisation and purpose limitation. A named privacy reviewer must decide what may be collected, joined, retained and displayed.

Do not force platform labels into one table

Two systems may use the same word for different events, identity rules, attribution windows or modelled values. Preserve the source label and definition. Create a common metric only when event, population, unit, period, deduplication, geography and exclusions match. Otherwise report separate columns and say that comparison is unavailable.

Language and market strata need the same care. Asset language can be recorded from the approved file. Participant language can be self-described within the research permission. Neither permits an analyst to classify a person's culture or comprehension from profile clues.

Compare attribution methods fairly

Descriptive monitoring reports observed events by source and time. Rule-based attribution assigns credit under a declared rule, such as a selected touchpoint within an agreed window. It supports reconciliation but does not establish what would have happened without the campaign.

Contribution analysis examines a theory of change and alternative explanations. It can strengthen or weaken a contribution claim but does not automatically estimate an effect size. Experimental or suitable quasi-experimental methods use a comparison that represents the counterfactual, subject to design assumptions.

The 2026 Magenta Book separates monitoring, process evaluation and impact evaluation. Its accompanying QPIE guidance says monitoring a change does not show that an intervention caused it and ties stronger causal confidence to a strong counterfactual. These are government evaluation records, not promises that a design will work in a commercial campaign.

Set decision rules without invented thresholds

The metric owner should state the action before viewing the result. Examples include investigate a data-quality flag, pause an inaccessible route, reconcile missing finance records or commission a stronger evaluation. Leave numerical thresholds blank until the buyer supports them with risk, capacity and evidence.

Claims made from the report also require review. CAP Code Section 3 requires documentary substantiation for objective claims and clear qualifications. A dashboard value alone is not substantiation if its population, definition or uncertainty is hidden.

Benchmark only comparable work

No authoritative England benchmark for multilingual social marketing performance was identified in this research. Do not manufacture one from broad UK advertising figures or global supplier studies.

A benchmark project must declare the target population, sampling frame, eligibility, recruitment, period, geography, language and market strata, metric cards, collection systems, missing-data treatment, weighting, uncertainty and publication rules. Participating organisations need equivalent definitions and auditable extracts. If comparability fails, publish "no usable benchmark" and keep the internal baseline separate.

Final reporting gate

Before circulation, a named statistical or evaluation reviewer should reproduce selected figures from frozen extracts. Translation and native-level reviewers inspect labels and notes; the subject reviewer checks meaning; an accessibility reviewer tests the final rendering. Privacy and PECR, advertising and consumer, security, employment, contract, finance and legal reviews remain independent.

Publish the cut-off, version, changes, limitations and next decision. Keep the report on editorial-draft-hold until those reviewers are named and every cited source is checked again on publication day.

Before you act

Common questions

What does a useful multilingual social marketing report actually contain?

It tells the reader what happened in a defined population and period, how the number was produced, what is missing and what decision follows. It does not turn delivery counts into audience comprehension, or platform attribution into causal impact.

When can two platform metrics be combined into one table?

Only when event, population, unit, period, deduplication, geography and exclusions match. Otherwise report separate columns and state that comparison is unavailable, because two systems may use the same word for different events, identity rules, attribution windows or modelled values.

How should language and market strata be handled in reporting?

Asset language can be recorded from the approved file, and participant language can be self-described within the research permission. Neither permits an analyst to classify a person's culture or comprehension from profile clues, names, locations or machine classifications.