The weekly dataset nobody wants to fund

Ask for budget to build a mix model and you'll usually get a hearing. Ask for budget to build the weekly dataset the model needs, and attention drifts. It doesn't sound like an outcome.

It is the outcome. Most measurement projects that disappoint didn't fail on the model. They failed on the data underneath it, months earlier, quietly.

What a modelling-ready dataset contains

At its simplest, one row per week (and per region, if you have regional data), with:

• Spend by channel, consistently defined, from every source: media agencies, platforms, invoices.
• Outcomes, both volume and value where possible: orders and revenue, applications and funded value.
• Price and promotions: when they changed and by how much.
• Competitor activity, even if it's only an index.
• Events: launches, outages, stock problems, strikes, anything that moved sales for reasons unrelated to marketing.
• Enough history: two years of weeks or more, ideally.

Why it's harder than it sounds

Spend lives in five systems with five definitions. Outcomes are counted differently by finance and marketing. Promotions are in someone's spreadsheet. Nobody wrote down the week the website broke. Each gap is small; together they decide whether a model can separate marketing from everything else.

Why it's worth funding first

• Every later model gets cheaper. Most of the cost of a first mix model is data work. Build it once, properly, and refreshes become routine.
• Results become defensible. When finance asks where a number came from, you can show them.
• Problems surface early. A good dataset reveals the channels that barely vary, or always move together, before anyone pays for a model that can't separate them.
• It serves more than one purpose. The same weekly table feeds dashboards, forecasts and experiment design.

How to make the case

Frame it as risk reduction, not infrastructure. The question for the budget holder is simple: do you want the next set of measurement numbers to be ones you can defend?

Nobody asks to see the dataset in the board meeting. It's still the thing deciding whether the numbers on the slide are true.