Prepare your data

Most of the work in a Marketing Mix Model happens before the model runs. The model is only as good as the table you feed it, and this is the step where runs most often go wrong. This page covers what your file needs, how to map your columns, and how to read the data check.

Setup does not have to happen in one sitting. Once you pick a file, everything you set on the page saves automatically as a draft, and the MMM Center holds it until you come back to resume or delete it.

What your data needs

You assemble one CSV yourself. Skeleton Key does not pull from your ad platforms automatically, so you export from Google, Meta, your analytics, and anywhere else, and combine them into a single table.

The shape is one row per time period, with your variables as columns:

• One date column, at a regular cadence. Weekly is the cadence Meridian is built around. Keep the gap between rows constant: every row a week, or every row a day, not a mix.
• One outcome column (your KPI). This is what the model explains, like revenue, orders, or signups.
• At least one channel of spend. One column per channel, holding what you spent that period.
• Enough history. As a rule of thumb the model wants roughly 15 periods of data for every thing it has to estimate, so more channels and more controls need more history. The setup flags anything under a year of weekly data as an error; Google's guidance for Meridian is closer to three years for a national model, or two when your data is split by region. The data check (below) tells you where you stand before you commit to a run.

Channel and column names become the labels you see in the results, so keep them readable (paid_search, linear_tv).

Rules the file must follow

The data check separates hard failures from judgment calls. The rules below are the hard kind: a file that breaks any of them fails inside the model engine every time, so the setup page stops the run and points at what to fix instead of starting a paid job that cannot succeed.

• Dates in YYYY-MM-DD format (2024-01-29, not 01/29/2024). The engine reads dates in exactly this shape and rejects anything else, including formats that look fine in a spreadsheet.
• No gaps, no mixed cadence. Every period present, all with the same spacing. A missing week fails the run, and so does a file that switches from weekly to daily rows partway through.
• Each date appears once. In a multi-region file, once per region and date.
• Every region covers the same dates. A region missing even one period that the others have fails the run.
• No blank cells in mapped columns. Fill gaps with a real value, or 0 where 0 is the truth (a channel that spent nothing that week).
• No negative values in spend, your KPI, or exposure metrics like impressions and reach. If refunds push a week's revenue negative, resolve that upstream before you upload.
• Complete reach and frequency sets. An R&F channel needs all three columns (reach, frequency, and spend); an organic R&F channel needs both reach and frequency. A partial set fails.
• Population on every row. In a multi-region file, the population column carries the region's population as a positive number, repeated identically on every row of that region.

After you upload, each column is assigned a role in the Column Mapping and Exploratory Data Analysis card. If your columns follow a naming convention Skeleton Key recognizes (see the tip below), many roles arrive already mapped, so you confirm or adjust rather than start from scratch. The roles are:

• Date and KPI (your outcome). For the KPI you also pick its type: Revenue ($) or A count (orders, installs, etc.). For a count, a value field appears where you can put a dollar figure on one outcome, like an average order value. Fill it and returns come back in dollars; leave it blank and they come back as impact per dollar instead. Read your results shows what each choice looks like.
• Paid media (spend, plus an optional exposure metric). The spend column is required; you can attach one exposure metric (impressions, clicks, or GRPs) to describe how the spend showed up.
• Paid R&F media (reach, frequency, and spend) for channels where you have true reach and frequency data, like YouTube.
• Organic (a metric only, no spend), for unpaid drivers like email sends or organic social.
• Organic R&F (reach and frequency, no spend).
• Controls, for outside conditions that move your KPI but aren't yours to steer: holidays, weather, competitor activity.
• Treatments, for non-spend levers you pull yourself, like a price change or a promo flag.
• Geo & population, if your data covers more than one region (see below).
• Anything you don't map stays Unmapped and is left out of the model.

Controls versus treatments

Both describe non-marketing influences, and the split trips people up. The question is not whether the column is a flag or a curve; it is whether the lever is yours. A treatment is something you decide and could have done differently, with a natural baseline to compare against: your price (baseline: the regular price), your promos (baseline: no promo). A control is an outside condition you can't intervene on: a holiday, the weather, a competitor's activity. So a price change or a promo flag belongs in Treatments, while a holiday flag or a weather series belongs in Controls. Smooth, recurring seasonality usually needs no column at all; the model's baseline absorbs it.

Variables to leave out

One exception is worth knowing. Search query volume looks like a mediator, and for most channels it is: TV and social make people search, so query volume takes some of their credit. But it also drives your paid-search results directly, since more searching means more auctions for your ads to win. Google's guidance for Meridian is to include query volume as a control when paid search is one of your channels, because it corrects a known bias that otherwise flatters paid-search returns. Map it under Controls, and expect your demand-driving channels to give up a little credit in exchange for a more honest paid-search number.

Multi-region, reach and frequency, and organic

Three optional inputs make a model stronger when your data supports them:

• Multiple regions. Add a geo column and a population column (the region's population, repeated on every row of that region). Every region is another set of evidence, so a geo model usually gives tighter estimates than a national one on the same history. It isn't automatic: a long, well-controlled national series can beat a thin panel of a few regions, so judge the result by its health verdict, not by which type you ran. Population is required so the model can scale regions of different sizes against each other. With national totals only, leave geo unmapped and run national.
• Reach and frequency. Channels mapped as Paid R&F media let the model find an optimal frequency rather than treating every impression the same. Use this only for channels where you genuinely have reach and frequency figures.
• Organic drivers. Unpaid drivers go in Organic (or Organic R&F) so their contribution shows up in the breakdown without being treated as paid spend. Because they carry no spend, organic channels get a contribution figure but no return on spend and no response curve.

The data check before you run

Before you fit a model, run the optional Exploratory Data Analysis over your mapped data. It returns one of three verdicts:

• Good — no statistical issues found, ready for a run.
• Caveats — some checks want a second look first; the data is usable with caveats.
• Issues — it flagged data problems worth fixing.

Behind that verdict, the check runs a battery grouped into a few sections:

• Spend and data volume — how your budget splits across channels and whether you have enough history for the model. This is where data points per parameter lives, the single most useful number: Meridian's rough guide is around 15, and if you're under it the page lists the levers in order of impact (add more history, combine or cut tiny channels, drop controls you can't defend, reduce knots). Channels with a tiny share of spend are shown here too, since they're hard to measure.
• Individual variables — whether each column moves enough for the model to learn from it (a near-constant column gives no signal), and which periods look like outliers, whether from a real event like a holiday or a genuine data error worth fixing at the source.
• Relationships between variables — how strongly your columns move together. Two that move in near lockstep are hard for any model to tell apart, so it would split the credit between them; the check flags the pair to combine or drop.
• Population scaling (multi-region runs only) — checks specific to geo models, where each region's population is used to weigh regions of different sizes against each other.

The data-check page also has a Download menu with three formats: a PDF of the report as you see it, a Markdown file of the findings, and the Technical report, Meridian's own EDA output as a standalone HTML file. The technical report carries detail the summary view leaves out, like the full list of flagged outlier periods.

The analysis opens as its own report page. Once it looks reasonable, head back to the setup page and start the run. Reading what comes back is the next page.