Working out the return on multilingual social spend with blank formulas and stated doubt

When a financial ratio can be filled in

A financial ratio is supportable only when the measured outcome, estimated counterfactual, approved contribution and full economic cost refer to the same population, period and decision. If any input cannot be aligned, leave the expression unfilled; a completed calculation would imply more certainty than the records permit.

Build cost from every relevant layer: licence and usage, translator or localiser labour, native-level and subject review, creative adaptation, accessibility, research, media, publishing, assurance, privacy, security, contract work, remediation and exit. Cash paid may not equal full economic cost, and recoverability does not remove effects on cash timing.

Keep gross revenue or service value, variable cost, accounting contribution, cash movement and wider non-monetary outcomes in separate fields. The finance owner should state the recognition basis and source. Do not put a monetary value on awareness, language inclusion, accessibility or avoided risk merely to complete a ratio.

Show VAT separately and retain evidence for each budget line using the budget evidence guide. Contract controls can identify exit, remediation and other contract work. Keep these amounts distinct from routine cash paid when reconciling the complete cost.

Keep each variable open until its currency, population, period and uncertainty are stated. The calculation starts by estimating the outcome above the counterfactual, then applies an approved contribution, before comparing that contribution with the full cost. These blank expressions make those steps visible without implying that a campaign result is known.

incremental outcome = observed eligible outcome - estimated counterfactual outcome

incremental contribution = incremental outcome x approved contribution per outcome

ROI = (incremental contribution - full economic cost) / full economic cost

For the England illustration, 1,000 approved orders sit against an estimated counterfactual of 850, with approved contribution of £12 per order. That gives 150 incremental orders and £1,800 incremental contribution. Set against full economic cost of £1,200, the illustrative ROI is 0.5, or 50 per cent.

Changing the estimated counterfactual to 900 moves the illustration's ROI to 0.0; setting it at 800 moves the result to 1.0. This span is an illustration, not a benchmark or a general expected result. It covers only those stated alternatives and does not capture uncertainty beyond them.

The same illustration split by language and channel

Take a UK retailer running one campaign in English and Welsh on Instagram and TikTok. All figures are invented to show the method. The £1,200 full economic cost is made of Instagram media £450, TikTok media £150, translation £250, native-speaker review £150 and publishing and assurance £200.

Split the 150 incremental orders by language: 100 English and 50 Welsh. At £12 each, incremental contribution is £1,200 for English and £600 for Welsh. Shared media and publishing costs are allocated by share of orders, 70 per cent English and 30 per cent Welsh, while translation and review are charged to Welsh.

That puts English cost at £560 and Welsh cost at £640. English ROI is (£1,200 - £560) / £560, about 114 per cent. Welsh ROI is (£600 - £640) / £640, about minus 6 per cent. Together they return the 50 per cent above.

The Welsh result is a monetary reading only. Inclusion and service to Welsh-speaking audiences are not given a value here. Report platform spend and outcomes separately where tracking allows, and say plainly where it does not.

For every input, record currency, population, period, evidence date, calculation owner, inclusions, exclusions and uncertainty. Do not calculate ROI if cost is zero, negative, disputed or incomplete, or if it uses another time basis. Publish a range only where each boundary has defensible evidence and its derivation is explained.

What a causal estimate can establish

A causal estimate needs a credible account of the outcome that would have occurred without the intervention. Before-and-after change cannot by itself separate campaign influence from seasonality, platform changes, price changes or other marketing. Matching may leave differences unmeasured, while random allocation may be unsuitable, underpowered or impractical.

Agree the design with an evaluator before launch. The Office for Statistics Regulation's administrative-data guidance concerns reconciling platform records with buyer records. It does not provide a result for an England campaign; treat it as guidance on quality assurance, not as evidence of campaign impact.

A strong causal claim is rarely warranted without a credible counterfactual. If the design cannot estimate outcomes in the absence of the intervention, keep conclusions within the evidence available. A cautious estimate of incremental outcomes is more defensible than describing an observed change as proof that the campaign caused it.

What an attribution rule assigns

Rule-based attribution assigns recorded events according to a stated convention; it does not estimate what would have happened without the campaign. Keep that allocation separate from causal evaluation. Use attribution methods to describe how the selected method assigns credit and what that allocation cannot show.

Record the tag, identity method, attribution window, duplicate handling, cross-device limitation, consent condition and later adjustments. Reconcile platform events with buyer records before relying on assigned results, and retain the source data and calculation owner. Those details make the convention inspectable without turning it into a causal estimate.

What monitoring records can show

Monitoring describes what was recorded, not what the campaign caused. For each measure, specify the event, eligible population, unit or denominator, timezone, reporting period, source query, exclusions and accountable owner. These details help readers interpret a recorded observation without presenting it as proof of impact.

Give each result a metric card. State the attribution convention and version, how missing data were handled, and any reconciliation with buyer records. Different platform and buyer totals may reflect definitions or collection problems; that gap alone does not show that the campaign changed an outcome. See measurement and reporting for the distinction between measurement and attribution.

When uncertainty requires a pause

Set stop conditions before reviewing results. If an issue weakens the evidence or approval, name who investigates, what evidence permits a restart, and whether earlier observations remain usable. Agree responsibilities in advance, rather than choosing a response after seeing the ratio or deciding which interpretation is convenient.

Pause if tracking consent changes, an event definition changes on the platform, target and comparison groups receive different offers, language versions make different claims, a material accessibility defect appears, or the finance owner cannot verify contribution. Where relevant, consult the ICO's data protection impact assessment guidance. Record the issue and what evidence would support resuming activity.

Place sensitivity checks beside the result. Vary one uncertain input at a time using an evidence-based alternative, including counterfactual outcome, contribution per outcome, missing data treatment and full cost. Show where the sign or decision changes. Do not choose only assumptions that produce a positive ratio, and explain the evidence behind each alternative.

CAP rules on misleading advertising require documentary support for objective claims before publication and set conditions for comparisons. Support for an advertising claim does not validate the analysis beneath it. A translated ROI claim should retain its qualifications and evidence, while the underlying calculation is checked separately.

If no credible causal design exists, report monitored and attributed observations under their precise labels, alongside limitations, sensitivity findings and a review date. A narrower statement with visible uncertainty is more useful than a precise ratio based on an invented effect. Tie any restart decision to the evidence and responsibilities agreed before review.

For consistent statistical quality, retain the definitions and evidence that let another reviewer follow each calculation. That includes the source data, the comparison groups used, the treatment of missing data and the assumptions that shaped a range. This record helps keep an objective claim distinct from an observed result.