Channels that move together can't be told apart

Most marketing plans move channels together. The big campaign launches, so TV, video, social and search all go up in the same weeks. The quiet period arrives, and they all come down. It makes sense for planning.

For measurement, it creates a problem no model can solve on its own.

Why the model gets stuck

A mix model works out what each channel did by watching what happens when that channel's spend changes and the others don't. If TV and social always rise and fall in the same weeks, there are no weeks where one moved and the other didn't. The data shows what they did together, but not what either did alone.

Statisticians call this collinearity. In plain terms: the model can't tell them apart.

How it shows up

The model will still split the credit between the two channels, but the split is fragile:

• Small changes move a lot of credit. Add a few weeks of data or change one setting, and the credit shifts between the two.
• Ranges are wide for both channels.
• Different models disagree about which channel did more, while agreeing on the total.

If you see these signs, the channels are probably tangled.

How to spot it before you model

Look at the weekly spend of each pair of channels over the last two years. If two lines rise and fall almost in step, expect trouble. A simple correlation check across channel pairs flags the worst cases in minutes.

Three ways out

1. Combine them. If they're always planned together, measure them as one, such as "video and social". You lose the split, but you get a reliable combined answer.
2. Bring in outside evidence. A geo test on one of the channels gives the model information the history can't. Used as calibration, it anchors one channel, and the other falls into place.
3. Change how you plan. Stagger flights, vary one channel while holding the other steady, or vary them differently by region. Within a few months, the data starts to separate them.

The honest answer in the meantime

Until then, report the pair's combined contribution with confidence, and the split with a clear caveat. If two channels always move together, the data can tell you what they did together, but not what either did alone. Saying so is better than a precise split nobody can defend.