The budget and the headcount it once demanded are no longer the barrier. Your data is. This guide shows what the method needs, lets you check where you stand, and tells you what to use if you're not there yet.
Choose how much detail you want. You can switch at any point.
Everything you need to judge whether mix modelling fits your team.
Mix modelling has a data threshold, not a company-size threshold. Five things decide whether it will work for you.
The model has to see each season more than once to tell a holiday rush from a campaign. One year is workable but fragile. Alongside the numbers it needs the dates of promotions, price changes and launches, or it will credit their sales to marketing.
Technical note. Two years of weekly data is about 104 observations. Each channel needs several parameters (effect size, carryover, saturation) on top of seasonality and controls, and a common rule of thumb is around ten observations per parameter. Pooling across regions or stores multiplies observations and is the main way shorter histories become usable. Daily data can help direct-response businesses, but it adds noise faster than it adds information and it doesn't add more seasons.The model learns by watching what sales did when spend changed. A channel held at the same budget every month is invisible to it, however large that budget is. Small teams often have an advantage here, because one person can change a budget in an afternoon.
Technical note. Identification comes from variance in each channel's spend that is independent of the other channels and of seasonality. Channels that scale together are collinear: the model can estimate their combined effect, but the split between them is then driven mostly by priors.Mix modelling exists to compare channels. With one or two, a holdout test gives a cleaner answer for less effort. With many, each extra channel needs more history to separate, so small ones are best grouped.
Weekly sales bounce around on their own. Marketing's contribution has to stand out from that bounce. That depends on how much you spend relative to your sales and on how steady those sales are, which is the real reason spend level matters.
Technical note. Weekly counts carry sampling noise of roughly one over the square root of the count: about 30% at 10 sales a week and 10% at 100. A channel driving a few percent of sales is hard to detect inside that without long history, pooled data or priors from experiments.The output is a recommendation to move money. It needs one person with a few hours a month to read it, and either the authority to act or the ear of someone who has it.
Seven questions, scored against the five requirements. Your profile updates as you answer.
These thresholds are rules of thumb. Published guidance on history ranges from one to three years, and guidance on spend varies widely.
Technical note. Region-level data, informative priors from experiments, and daily data in direct-response businesses can all relax these thresholds. Treat the result as a screening check and not a substitute for inspecting the data.A mix model involves five jobs. Software has taken over the three that used to need statisticians. The two that remain need knowledge of your business.
Bringing spend from every ad platform and sales from your store or CRM into one weekly table.
Connectors in a measurement platform, or a spreadsheet kept by hand.Choosing the statistical form, fitting it and tuning it.
Software. This was the statisticians' job.Testing whether it can predict weeks it hasn't seen, and whether its answers survive small changes.
Software runs the checks. You confirm they passed.Turning contribution and return figures into “this channel has room, that one is flattening”.
You. It takes knowledge of the business, not statistics.Moving budget, then watching what happens.
You, or whoever holds the budget.Software removes the statistics. It doesn't remove judgement. Get expert help when results contradict a test or plain common sense, when your business is unusual (a few large deals, extreme seasonality, most sales through a marketplace), or when the decision is too large to reverse.
You are the owner, the analyst and the decision-maker, so keep the model simple enough for one person to run.
For most small teams this adds up to a few hours.
Once a quarter, add one incrementality test on your largest or least certain channel, and use it to check the model.
There are three ways to get a mix model. Each asks you to pay in a different mix of money, time and skills.
Measurement earns its place when the money it helps you move is larger than what it costs.
As a rough guide, take a tenth of your annual ad spend as the amount better allocation might shift. If a year of measurement would cost close to that or more, wait, and use the free tools first: platform reports read side by side, one neutral attribution view, and a single test.
Being short of the threshold isn't a reason to measure nothing. Two other methods work with far less data, and the third step makes sure the wait is useful.
For weekly decisions
Count every platform by the same rules in one place, and use the result to steer campaigns inside each channel. It works from the first day you have tracking, and it needs no history.
For the big question
Hold ads back from some regions or audiences for a few weeks and compare sales. One test on your largest channel tells you whether it earns its budget, and the pause itself is variation a future model can learn from.
So the wait counts
For how the three methods fit together, read the Guide to Measurement.
Bconic offers a free 30-minute audit of your current measurement.
