Media Mix Model

A Media Mix Model measures how much each of your channels actually contributed to sales, signups, or whatever outcome you care about. It reads your historical spend and results, separates what your marketing drove from what would have happened anyway, and gives you a return figure for each channel.

This is a different question from the one your attribution dashboard answers. Attribution assigns credit along individual user journeys (which click came last, which touchpoint gets the conversion). A Media Mix Model never looks at individuals. It works from aggregate, period-by-period data and estimates the incremental lift each channel caused, including the channels that attribution can't see well, like TV, out-of-home, or upper-funnel video. It also accounts for things outside your marketing (seasonality, price, promotions) so they don't get miscredited to a channel.

The Skeleton Key MMM currently runs on Meridian, Google's open-source modeling engine, but is built to support other open-source MMM models over time.

What Skeleton Key adds

Running Meridian by hand means a Python and GPU stack, hand-coded data inputs, choosing the right statistical assumptions, and reading raw diagnostics. Skeleton Key handles all of that, so a media planner or marketing manager can run an MMM without writing code. On top of the raw model, it adds:

• Guided setup. You bring one CSV. Skeleton Key recognizes and maps your columns, enforces the model's data rules so you can't start an invalid run, and provisions the compute for you.
• Plain-English insight. Every chart comes with an explanation, a glossary translates the statistics, the data check flags problems before you run, and multi-region runs get per-region performance views. You can download Meridian's full report for both the data check and the results.
• A link to your strategy. When you run a Strategy, the MMM results inform the Goals Strategist and Media Strategist, so your OKRs and media mix are grounded in what the model measured, not assumptions.

What it produces

When a run finishes, you get:

• Return by channel. What each channel returned on its spend, shown as a range rather than a single number, because the model is honest about its own uncertainty.
• Diminishing returns curves. How the next dollar in each channel performs as you spend more, and where a channel starts to saturate.
• What drove your results. A breakdown of your outcome into the baseline (what you'd have earned with no marketing) and the share each channel contributed.
• Optimal frequency. For reach and frequency channels, the weekly frequency that returns the most per impression. This only appears when you map reach/frequency channels.
• A model-health verdict. A plain read on whether the model settled and fit your data well enough to trust the numbers, scored out of 100 with a verbal verdict.

Paid and organic are measured differently

Paid channels are the ones you scale with money: budget up, budget down, add a channel, cut one. For those the model gives you the full toolkit, including return on spend and the diminishing-returns curves.

Organic drivers (email sends, organic social, newsletter reach) carry no spend, so there is no return to compute and no spend curve to draw. The model still measures their contribution and shows it in the breakdown, which keeps their credit from being wrongly handed to your paid channels. They are also harder to steer: you control how many emails you send, but not how far a post travels.

When to use it

A Media Mix Model answers strategic budget questions ("which channels are worth more money," "where are we hitting diminishing returns") rather than tactical ones ("did this creative beat that one"). For tactical questions, an incrementality test is the better tool. A Media Mix Model is an aide to judgment, not a replacement for it.

One caveat to hold from the start: treat the numbers as directional. They tell you the relative story (this channel is working harder than that one, this one is saturating) more reliably than they pin down an exact return to the second decimal. The reading results page explains how to read the model-health verdict so you know how much weight the numbers can carry.

This is the same class of model that measurement consultancies build, run in-house on your own data instead of outsourced. It is in beta.

Do you have the data for it?

Three things decide whether an MMM can work for you. Check them before you assemble a file.

• Enough history. A year of weekly data is the floor the setup enforces. Google's guidance for Meridian is closer to three years for a national model, or two when your data is split by region, because every region is another set of evidence. Under that, expect directional reads rather than confident ones.
• Varying channel spend. The model learns by watching your results change when your spend changes. A channel that spent the same amount every week gives it nothing to learn from, and the data check flags it. Flighting, seasonal pushes, tests, and budget shifts all count as useful variation.
• Meaningful KPI contribution. The model splits your KPI into a baseline (what happens without marketing) and the share your channels drove. If most of your KPI comes from organic demand rather than the channels you pay for, the model may find little or weak signal from your paid channels, and the returns it reports will be wide and uncertain.

How a run works

Two pages walk through the halves of that flow that need the most care. Prepare your data covers getting your data in and mapping your columns. Read your results covers reading the run once it finishes.

Where to find it

Click MMM in the left sidebar to open the MMM Center, where your past runs live. Start a new one with New MMM run, or open Try a sample run to see finished results on example data before you bring your own.

Limits and considerations

• It needs the data described above. Short histories or channels with little spend variation produce weaker, less trustworthy runs.
• The model runs on sensible defaults. The advanced run settings (the number of sampling chains and draws, the ROI prior, and how much the baseline can bend) are there for experienced users, and changing them without a reason, the ROI prior especially, can bias the results.
• It works best alongside incrementality experiments, not instead of them. If you've run a lift or incrementality test, you can set the model's ROI prior from that result, which calibrates the model toward a return you measured directly. The two together beat either alone.