A delayed advertising response changes how finance reads weekly returns

A campaign can influence purchases after the week in which the advertising runs. Comparing only that week's revenue with that week's spend can therefore give an incomplete view of the assumed response. Finance needs a defined horizon and evidence about the delay before translating a marketing model into a budget decision.

Research by Jin and colleagues at Google develops media-mix models with carryover and diminishing-return effects. The paper treats those effects as model components to estimate, not universal constants. The simple example below uses an explicitly assumed geometric carryover rule to make the timing arithmetic visible.

Keep the actual spending schedule separate from the response model. The model may carry an influence into future periods, but the business did not necessarily spend the same money again in those periods. A statistical index should not be mistaken for a supplier invoice or a new cash commitment.

Let the modeled activity index equal current spend plus half of the previous week's index. Start with a zero index, spend $100,000 in week one and spend nothing in the following weeks. The index is then 100,000 in week one, 50,000 in week two, 25,000 in week three and 12,500 in week four.

These values are units of the chosen modeled index, even though the input was dollars. Actual campaign spending remains $100,000. Adding the first four index values gives 187,500, which represents accumulated modeled activity under this rule, not $187,500 of payments.

The indefinite sum of the index from that pulse is 200,000 because the remaining influence halves each week. That total depends on allowing the geometric sequence to continue. A model with a finite lag window or a different decay pattern would produce a different cumulative index.

Now assume, purely for illustration, that incremental revenue equals twice the activity index and that incremental contribution before campaign spend is 40% of that revenue. Modeled revenue in the first four weeks is $200,000, $100,000, $50,000 and $25,000.

Week one produces $80,000 of modeled contribution before the $100,000 spend, leaving negative $20,000 after that spend. Across four weeks, modeled revenue totals $375,000 and contribution totals $150,000, leaving positive $50,000 after the campaign cost. Under the indefinite geometric continuation, the modeled net contribution would reach $60,000.

Those figures are consequences of the assumed response coefficient and decay, not evidence that the campaign works. A model can produce an attractive delayed return if its coefficient is optimistic. Keep the measured evidence and uncertainty behind that coefficient beside the cumulative result.

Some carryover transformations normalize their weights so the total weight is one. Others, like the simple recursion above, do not. A coefficient estimated under one convention cannot be moved unchanged to the other without checking the scale. Otherwise the same historical spending can appear to produce a different total effect solely because the transformation changed.

The decay parameter also belongs to a time unit. A 50% weekly retention of influence is different from a 50% daily retention. Changing the dataset from weekly to daily while preserving the same parameter would change the modeled persistence dramatically.

Record the lag length, normalization, time frequency and any initial carryover balance. Starting a model at zero in the middle of an ongoing campaign can omit the influence of earlier spending. Preserve enough prior history or use a supported initialization assumption.

When comparing continued spending with stopping, include the existing campaign's remaining influence in both alternatives where appropriate. Some future response may occur even after new spending stops. Crediting that entire existing tail to the continuation decision would overstate the contribution created by its additional commitment.

A sale may occur promptly but reach the analytics report later. That is a measurement delay, not necessarily a delayed customer response to advertising. Review transaction dates, reporting dates and attribution-window behavior before interpreting the observed pattern as economic carryover.

Likewise, revenue recognition and cash collection can occur after the purchase event. A marketing-response model based on purchases needs additional steps before it becomes a revenue or cash forecast. Keep those transformations identifiable so the same lag is not applied twice.

Use dated cohorts or other suitable evidence to examine the timing. If a platform revises prior conversion totals, retain the earlier data vintage when reviewing the model's historical forecasts. A fit based on fully matured reports may have had more information than the weekly budget owner possessed at the time.

Advertising spend can rise before a seasonal sales peak even if the peak would occur without the spending. A lagged statistical relationship alone does not establish that advertising caused the later sales. Use an appropriate identification strategy, experimental evidence or other supported calibration for the intended decision.

A response estimated under one offer, audience or creative may change when those conditions change. A longer modeled tail should not become a permanent justification for a campaign whose measured economics deteriorate. Review which evidence supports persistence and which part remains a judgment.

Test the decision under plausible alternative decay and response assumptions. If the campaign only covers its cost when the effect lasts unusually long, that dependency deserves attention before the next commitment. Present the range without assigning unsupported probabilities to every scenario.

Place media bills, production costs and other campaign payments on their real dates. Then translate supported incremental purchases into contribution and collection timing. The business may need to fund the campaign before the modeled longer-term return can supply cash.

Retain the spending extract, response model, horizon and assumptions with the approval. Finance can then distinguish a short measurement window from a weak campaign, while making clear that a delayed modeled benefit remains an estimate until the underlying evidence supports it.