Media mix contribution for a physical channel

Marketing mix modelling estimates each channel's contribution to sales from historical variation in spend and outcomes. It works best for channels with substantial spend, meaningful variation over time, and a long history.

In-store media in a fragmented channel frequently has none of those, which limits what a mix model can say about it.

Why do small channels fit badly?

Because the model needs variation to attribute effect, and a channel that is a small share of spend contributes little variation to a national sales series. Its coefficient is then estimated imprecisely and can be unstable across model specifications.

That is not evidence the channel does not work. It is evidence that the model cannot see it clearly.

What makes it worse in this channel?

Geographic granularity. In-store media operates at store level and national mix models operate at market or national level, so the channel's variation is averaged away before the model sees it.

A model built at a finer geography, where the channel's store-level variation survives, is a considerably better fit.

What is the alternative?

A designed test. A holdout or matched-market experiment measures the channel directly and produces a causal estimate rather than a modelled coefficient.

For a channel with store-level activation, that option is available in a way it is not for broadcast, and it is a genuine advantage worth using.

Can test results be fed into a mix model?

Yes, and calibrating a model with experimental results is established practice. It anchors a coefficient the model would otherwise estimate poorly.

That is the most productive relationship between the two approaches for this channel: experiments inform the model rather than competing with it.

What should a brand not do?

Conclude from a small or unstable mix model coefficient that the channel does not work. An imprecise estimate is an absence of evidence rather than evidence of absence, and the distinction matters commercially.

Equally, a large coefficient on a small channel should be treated with caution rather than celebrated.

How should the channel be represented in a model?

With accurate spend and, ideally, with a delivery measure rather than a spend proxy. Delivered plays or store-weeks of activity describe the channel's actual presence better than invoiced spend does.

What about attribution models?

Digital multi-touch attribution does not extend to a physical channel with no identified exposure, and attempts to include it are usually assumption rather than measurement.

What is the honest position?

That a mix model may not resolve this channel, that a designed test can, and that the two together are stronger than either. Saying so is more credible than arguing for a coefficient the data cannot support.