MMM Build

A marketing mix model built to change next quarter’s budget, not to decorate a slide deck.

Who it’s for

Brands spending across several channels that need to know what each one really contributes, where returns are flattening, and how to split next year’s budget. Also brands whose existing model was built, presented and then ignored.

What you get

  • A bespoke model built on your data, using the right tool for your situation: PyMC, Meridian, Robyn or custom econometric code.
  • Channel contributions and the baseline, the sales that would happen without paid media, with uncertainty ranges.
  • Response curves showing where each channel’s returns start to diminish.
  • Calibration against experiments where you have them, so the model is checked against something real.
  • Budget scenarios showing what changes if you move money, and how confident we can be.
  • A handover so your team can read, challenge and re-run it.

Try it: change the assumptions

Search: half of the maximum response is reached at a spend of 60. Returns start to flatten around 53.

Illustrative

Effects linger

Carry-over (adstock): what is left of a one-off pulse in the weeks after it

Bars fall week by week after a one-off pulse of 100. A higher carry-over setting makes the effect last longer.

View data
Remaining effect of a pulse of 100, by week, for each channel’s starting carry-over
WeekSearch (0.3)Social (0.5)Video (0.7)
1100.0100.0100.0
230.050.070.0
39.025.049.0
42.712.534.3
50.86.324.0
60.23.116.8
70.11.611.8
80.00.88.2
90.00.45.8
100.00.24.0
110.00.12.8
120.00.02.0
Illustrative data. Built to show how the method works, not a client result.
Illustrative

Returns flatten

Diminishing returns (saturation): response as a share of the maximum, by spend

Three curves rise and flatten as spend increases. Channels with a lower half-saturation spend flatten sooner, so extra spend buys less.

View data
Response as a share of maximum, by spend index, at each channel’s starting settings
SpendSearchSocialVideo
00%0%0%
1010%10%2%
2021%18%6%
3030%25%11%
4038%31%16%
5045%36%21%
6050%40%26%
7055%44%31%
8059%47%35%
9062%50%39%
10065%53%43%
11067%55%47%
12070%57%50%
13072%59%53%
14073%61%56%
15075%63%58%
16076%64%61%
17078%65%63%
18079%67%65%
19080%68%67%
20081%69%68%
Illustrative: how carry-over and diminishing returns shape what an extra dollar buys. Illustrative data. Built to show how the method works, not a client result.
Illustrative

Where sales come from

% of sales. The baseline is separated from media.

A waterfall showing a 62% baseline and media channels adding up to 100% of sales.

View data
Contribution to sales (% of sales)
SourceContributionCumulative
Baseline (organic demand)62%62%
Search14%76%
Social9%85%
Video5%90%
Email and CRM4%94%
Brand halo3%97%
Other3%100%
Illustrative: where sales come from, with the baseline separated from media. Illustrative data. Built to show how the method works, not a client result.

How it works

  1. Data readiness check. What you have, what’s missing and what it limits.
  2. Model design. Structure and priors chosen with you, grounded in how your business actually runs.
  3. Estimation and validation. Out-of-sample tests, sensitivity checks, comparison with experiments.
  4. Decisions. Scenarios and a plan, not just coefficients.
  5. Handover and refresh. How to keep it alive after I step back.

Timelines depend on data readiness. The audit or the readiness tool will tell you what to expect.

What I need from you

Weekly or daily sales and spend history by channel (ideally two years or more), a record of promotions, price changes and other events, and a business owner who will use the answer.

Evidence you’ll reach

Level 2, modelled evidence, moving towards Level 3 to 5 when calibrated against experiments. I’ll tell you which level each finding sits on.

Evidence levelLevel 2, towards 3 to 5 when calibratedSee the Evidence Ladder
Do you build models that other tools can’t?

The tool matters less than the design, the validation and what you do with the result. I choose the tool that fits your data and question.

What if our data isn’t good enough?

Then I’ll say so, and tell you what to fix first. That is a better outcome than a confident model on weak data.