What data you need for marketing mix modelling.

A marketing mix model estimates what each channel contributed by watching how a client's outcomes moved when their spend moved. It needs no tracking, no cookies and no user-level data, which is why interest in it has returned as tracking has become less complete.

The catch is in the first sentence. If spend on a channel never changed, the model has nothing to learn from, and if the history is short, it cannot separate a channel's effect from the season it happened to run in.

Five things, and the first is the one most often got wrong.

• An outcome series

  Revenue, orders or leads from the client's own system of record, not the conversions each platform attributes to itself. The model is meant to decide the attribution, so it cannot be fed somebody else's.

• Spend per channel per period

  Usually weekly, with impressions or clicks alongside where the channel reports them. Every channel, including those with no reporting API.

• Enough history

  Enough to cover several full seasonal cycles, which in practice means years of data rather than months.

• Outside factors

  Price changes, promotions, holidays, stock problems, distribution changes. Anything else that moved sales, so the model does not credit it to whichever channel was spending at the time.

• Consistent geography

  If you model by region, spend and outcomes both have to be split the same way, for the whole history.

History that looked complete turns out to have holes: an account paused and reconnected, a channel renamed halfway through, a currency that changed, a conversion definition revised in the second year. Spend that never had an API sits in spreadsheets in three different formats.

Each of those has to be repaired before the model sees it, because the model cannot tell a gap from a pause, or a renamed channel from a new one. Teams who start with the model usually end up back at this step, and it is the slow one.

Platforms limit how far back their reporting interfaces will return data, so the history a model needs next year has to be collected this year. The cheapest decision in the whole exercise is to start keeping a client's daily spend and outcomes now, in a warehouse you control, even if a model is a year away. An agency that does this for every client can offer a model later without asking anyone to wait for the data.

A mix model answers how to split a budget across channels, slowly, from history. Attribution answers which campaign inside a channel is doing better, quickly, with each platform's own bias. Incrementality tests answer whether one channel really caused anything, at the cost of holding some spend back.

The three are better together than apart. A model is most reliable when tests have been used to check the channels it is least sure about.

SightShark does not build the model. It builds what the model reads: spend from 200+ sources and outcomes from each client's store or CRM, in one currency, on one grain, under one name per channel, accumulating daily in your own data warehouse. Manual uploads bring in the spend that has no API, so the history is complete rather than merely long.