This page is here so you can check our figures without asking us for anything. Every parameter below comes straight from the calculation engine: change it there and this table changes with it.
The audience is split into a test group and a control group. We treat conversions as counts. The logarithm of the ratio between the two groups has approximately a variance inversely proportional to the number of conversions, the number of weeks, and the product of the holdout share with its complement.
From that follows the number of measurement weeks you need to separate a given effect from chance. And in reverse, which is often more useful: for a given number of weeks it gives you the smallest effect you can still demonstrate.
We do not calculate with pure Poisson noise. Real weekly figures swing more than that, driven by season, promotions, stock and competitive pressure. So we multiply the variance by a factor that depends on your business model: the further the conversion sits from media exposure, the higher that factor.
On top of the measurement weeks comes a maturation period. With a long sales cycle, todayu2019s conversions belong to media from months ago, and that lag has to fit inside the window.
W is the number of weeks you need to measure before a relative lift delta can be told apart from chance. C is conversions per week across your whole population, h the share you hold back, phi the variance factor of your business model, and z_alpha and z_beta follow from your confidence level and your power.
This is the form that changes the conversation with finance. Not "how long does it take to demonstrate five per cent", but "what can I see at all within my own decision rhythm".
Within a decision cadence, measurement room is what remains once maturation is taken off. If maturation runs longer than the cadence, there is no window inside a cycle and we calculate across a full year. That is the sharpest finding this tool produces: anyone judged on a period shorter than their own sales cycle systematically measures the effects of a period they can no longer steer.
They appear below exactly as they sit in the calculation engine, not transcribed. Anyone who wants to dispute a value knows precisely which one.
| Choice | Value |
|---|---|
| Webshop or DTC | 1,6 |
| Subscription or SaaS | 1,8 |
| Lead generation B2C | 2,0 |
| Non-profit or public sector | 2,2 |
| Lead generation B2B | 2,6 |
| Retail with stores | 2,8 |
The further the conversion sits from media exposure, the more your weekly figures swing for reasons that have nothing to do with media. This is an estimated parameter, not a measured value; it is the first number we will calibrate once enough calculations have come in.
| Choice | Value |
|---|---|
| Less than a week | 1 |
| 1 to 4 weeks | 2 |
| 1 to 3 months | 6 |
| More than 3 months | 13 |
| Choice | Value |
|---|---|
| Under 250k | € 150.000 |
| 250k to 750k | € 500.000 |
| 750k to 2M | € 1.350.000 |
| 2M to 5M | € 3.500.000 |
| Above 5M | € 7.000.000 |
| Choice | Value |
|---|---|
| Less than 6 months | 20 |
| 6 to 12 months | 40 |
| 1 to 2 years | 78 |
| More than 2 years | 130 |
| Choice | Value |
|---|---|
| 1 to 3 | 2 |
| 4 to 6 | 5 |
| 7 to 10 | 8 |
| More than 10 | 13 |
| Choice | Value |
|---|---|
| Weekly | 4 |
| Monthly | 8 |
| Quarterly | 13 |
| Annually | 39 |
A model with k channels needs roughly 3k + 8 parameters: adstock, saturation and a coefficient per channel, plus a baseline. We test the number of weekly observations per parameter. Below three the model is not identifiable, between three and five it only works Bayesian with strong priors from experiments, and from five it is defensible. The binding condition is not budget but history.
Two answers always come back as not achievable, whatever the volume. Fully national buying, because then no control area exists: that is a buying problem, not a statistical one. And the language border, because Flanders and Wallonia differ structurally in media consumption, competitive pressure and purchasing power, so the gap you measure is not your campaign. This is deliberately hard, and it is where this tool differs from an international calculator.
A mark out of a hundred would suggest measurability is a report card. It is a choice with a price.
We express agency effort in PAR, not in euros per day. That is a deliberate limitation.
We say what is achievable, not what works. The second one needs your data.
The budget is not the problem, the number of observations is. Weekly data over two years gives you about a hundred data points. With eight channels plus adstock, saturation and coefficient parameters per channel you are already past thirty parameters. Below five observations per parameter the model is not identifiable and produces estimates driven entirely by the priors.
Flanders and Wallonia differ structurally in media consumption, competitive pressure, purchasing power distribution and seasonality. The gap you measure between them is therefore not attributable to your campaign. A valid geo experiment works with matched regions inside a single language region.
The number of measurement weeks follows from W = phi times the square of the sum of z-alpha and z-beta, divided by the product of weekly conversions, the holdout share times its complement, and the square of the natural logarithm of one plus the lift. At a thousand conversions per week, a ten percent holdout and a five percent lift to detect, the best case is thirty-seven weeks.
One page that marketing and finance sign together before the budget year starts. It fixes what gets measured, at what confidence, what explicitly does not get measured, which decision rests on which level of evidence, and how often it is reviewed. It moves the burden of proof discussion forward, to the point where it can still be negotiated.