Search spend responds to demand
Paid search often increases when more people are already searching. A weekly correlation between spend and sales can therefore reflect both advertising response and demand capture. Treating the spend series as an independent intervention without examining that feedback can inflate the apparent effect.
- Separate search types
- Map demand drivers
- Challenge control choices
- Test the marginal change
Separate branded and non-branded activity where the data supports meaningful estimation. Keep consistent definitions when campaigns or match types change. The confounding guide explains the central problem: a common driver can move both media and sales.
The response you infer depends on transformation choices and what the data can separate. The how marketing mix models work learning path connects this idea to the assumptions, practical checks and next steps.
To work through the connected decisions, read MMM channel grouping: when separate ROAS is unsupported and Email and MMM: use holdouts to separate selection from lift.
Draw the demand pathway before adding controls
A brand campaign can increase branded searches, which can increase paid search clicks and sales. Controlling for branded search volume may remove part of the brand campaign's total effect. On the other hand, an external demand event can raise both searches and sales, making a demand control valuable.
The same observed variable can play different causal roles. Write down which effect the model intends to estimate and use domain knowledge to decide whether a control is upstream or downstream of the intervention. Do not select controls solely because they improve fit.
Use an illustrative counterfactual
Suppose paid search reports £100,000 of attributed revenue on £20,000 of spend. A credible experiment for the relevant population estimates that £30,000 of revenue would disappear without the intervention. Attributed ROAS is 5, while the experiment's incremental revenue ROAS is 1.5. They answer different questions, and neither number should be relabelled as the other.
The experimental result applies to its tested treatment, time window and audience. A brand-keyword pause is not automatically evidence for all non-brand search. Preserve the definition when using the result to challenge or calibrate an MMM.
Diagnose the proposed budget move
Inspect auction price changes, impression share constraints, query mix and major website changes. More spend can purchase the same exposure at a higher price rather than expand exposure. Compare the model's spend response with what the delivery data makes plausible.
If the allocator recommends a large search increase, ask which additional searches the budget can buy and whether they have similar intent. Constrain the move to a credible range and measure the outcome. Use the incrementality testing guide to design an intervention that answers the next-pound question.
A practical review pack contains a brand/non-brand definition, a causal sketch, demand and auction diagnostics, the experiment scope and a bounded recommendation. High predictive accuracy alone does not resolve demand endogeneity, so report sensitivity to plausible control choices.
Sources and further practice
This article’s worked example is illustrative, not a measured client result. Use the linked methodology to check the assumptions before applying it.