Social media measurement works better when the question comes before the dashboard. Are we trying to establish whether a campaign increased purchases, whether a message changed brand perceptions, or whether the programme produced a financial return? Those questions need related but different evidence.
Substantively reviewed and updated: 22 October 2026. This article replaces fixed holdout and time-series rules with a measurement process based on the decision, available data and uncertainty.
Define return before choosing a method
Likes, reach, registrations and brand lift may be useful outcomes, but they are not themselves financial ROI. Platform-attributed conversions also differ from incremental conversions: some purchasers would have bought without the campaign.
For an explicit contribution-based calculation, define incremental contribution before the programme cost being assessed as incremental revenue minus the associated delivery and service costs. Then:
ROI = (incremental contribution before programme cost − incremental programme cost) / incremental programme cost
State the horizon and include the relevant media, creative, agency, tools and staff costs in programme cost. Avoid deducting the same cost in both contribution and programme cost. For a decision about expanding an existing programme, assess the additional costs and outcomes of that expansion. A full-programme evaluation may use a broader cost scope; explain the choice.
A hypothetical financial example
Suppose a suitable experiment estimates $30,000 of incremental revenue over the measurement horizon. Assume associated fulfilment and service costs are $12,000, leaving $18,000 of incremental contribution before programme cost. If the programme costs $10,000, its estimated net contribution is $8,000 and ROI is 80%: ($18,000 − $10,000) / $10,000.
Revenue divided by media spend would answer a different question and omit costs. The 80% result is also an estimate, not certainty: report uncertainty in lift and test how plausible changes in margins or delayed outcomes affect the decision. Do not extrapolate repeat purchases indefinitely without retention evidence.
Incrementality experiments
A randomised holdout compares outcomes between units assigned to treatment and control. Depending on the channel and delivery system, units might be eligible people, households or geographic markets. Analyse the effect of assignment under the design; comparing people who happened to see an ad with those who did not can introduce selection bias.
Choose the allocation and duration using baseline variation, the minimum effect worth detecting, desired statistical power and business constraints. A 10% spend holdout is not automatically statistically sufficient. Consider the number of independent units, geographic differences, exposure spillover, lagged outcomes and whether the campaign can actually be withheld.
Predefine the main outcome, exclusions and analysis before reviewing results. Check that assignment and measurement worked, and report uncertainty and practical importance as well as statistical significance. A non-significant finding may be inconclusive rather than evidence of zero effect.
Time-series counterfactual analysis
When randomisation is unavailable, a time-series model can estimate what might have happened without an intervention. Google’s CausalImpact package uses Bayesian structural time-series models to estimate this counterfactual and the difference from observed outcomes. Its original research paper explains the approach.
The controls must not be affected by the intervention, and their relationship with the treated series must remain sufficiently stable after it. Simultaneous pricing changes, distribution shifts or other campaigns may undermine the interpretation. Report posterior credible intervals for this Bayesian implementation, rather than describing them as frequentist confidence intervals.
There is no universal rule that 50 weekly observations makes an analysis adequate. The pre-period must support the relevant seasonality and relationships; assess fit, placebo periods and sensitivity to controls and modelling choices. A plausible-looking graph alone does not establish causality.
Brand lift and creative testing
Brand studies can measure changes in awareness, message association or purchase intent. State who was sampled and how treatment and control were formed. Survey response bias and differences in exposure can affect results. Treat lift in an attitude as an outcome in its own right; converting it to sales or ROI requires additional evidence.
Randomised factorial experiments vary combinations of creative or audience factors and can estimate interaction effects. They differ from observational regression with multiple variables. NIST’s experimental-design guidance explains why changing one factor at a time can miss interactions. More combinations also demand a suitable design and enough information; multivariate testing is not inherently more rigorous in every situation.
Bring the evidence together
I would keep a short measurement plan for each decision: the outcome, horizon, causal question, method, assumptions, cost definition and action threshold. Use experiments where feasible, modelling where its assumptions are defensible, and brand research for the perceptions it actually measures. Report disagreements instead of averaging unlike measures into a single reassuring number.
For broader channel-budget questions, see the MMM primer. For an affordability ceiling before acquisition, see the proposed DMA-CPA heuristic. Neither replaces evidence of incremental financial return.