Measure · Media mix model

The next dollar, not the average one.

A Bayesian media mix model fit on your weekly spend and the revenue you booked. It shows what each channel returned, and what its next dollar will.

Paid search

Modeled weekly revenue by weekly spend, with the 80% credible band

Paid search
$0$0k ($0k to $12.2k)
2$29.1k ($19.3k to $38.9k)
3$51.1k ($43.7k to $58.6k)
4$67.8k ($62.8k to $72.9k)
$20k$80.5k ($76.9k to $84.1k)
6$90.1k ($86.5k to $93.7k)
7$97.3k ($93.7k to $100.9k)
8$102.8k ($99.2k to $106.4k)
$40k$107k ($101.2k to $112.8k)
10$110.1k ($99k to $121.3k)
11$112.5k ($96k to $129.1k)
12$114.3k ($92.4k to $136.3k)
$60k / week$115.7k ($88.4k to $143.1k)
Sample company data.

Next dollar returns

Next dollar returns
Microsoft Ads1.8×
Paid social1.4×
Paid search1.3×
Sample company data. The largest channel is not where the next dollar earns most.

What a fit tells you

Six numbers behind every budget.

Each one is estimated from your own weekly rows, not assumed from a benchmark.

  1. Baseline

    The revenue you would book with media at zero: seasonality, brand and organic demand. Channels are measured against what's left.

  2. Contribution

    What each channel was worth last quarter, with a range finance can see.

  3. Carryover

    How long a dollar keeps working after the week it's spent, and how fast it fades.

  4. Saturation

    Where each channel's curve flattens. The reason doubling a budget rarely doubles revenue.

  5. Marginal return

    What the next dollar returns in each channel at today's spend, ranked.

  6. Range

    An 80% credible interval on return, so you know when a number is ready to act on.

What it needs

What a fit needs from your data.

The data is checked before any fit runs. When it falls short, you see why and what to add.

Weekly history

52 weeks

The least it will fit on. 104 weeks is recommended, so seasonality and trend can be told apart.

Spend variation

15%

A channel whose weekly spend barely moves is flagged. A budget that never changes gives the model nothing to compare.

Channel history

26 weeks

A channel active for fewer weeks gets a wide range. Grouping it with a similar channel is the usual fix.

Before you see it

Every fit is checked first.

When a fit falls short, you get the reason and the fix, never a sample number in its place.

  1. Reconciles to booked revenue

    Channel contribution can never add up to more than the revenue you booked.

  2. Converges

    The sampler's chains have to agree before a figure is shown.

  3. Forecasts weeks it never saw

    With more than 60 weeks of history, a fit has to predict eight recent weeks held out of training.

One fit, everywhere

The same curves run the rest of the product.

Find where the next dollar earns most.

Data refreshes nightly. The model refits every week.