CMO playbook & advanced tactics for marketing mix modeling

Two numbers for the same quarter and the same campaign. Meta's dashboard credits it with something like three times the conversions your mix model gives it. Neither figure is a mistake; they answer different questions from different data. One of them will set next year's budget, and that arbitration lands on your desk, not your analyst's.

Model-led budgeting at the top is a claim on money, which is why every channel owner whose number gets cut will attack it. If finance does not believe the model, it changes nothing. The plan reverts to last year plus inflation, and the modelling invoice becomes a line item you defend twice.

The arbitration you cannot delegate

A platform report sees only its own touchpoints, inside its own attributionattributionA framework for assigning credit to the touchpoints that contributed to a conversion, so you can measure which channels and interactions actually drive results.View full definition → window (Meta's default is 7-day click and 1-day view), and it has no view of your price, your distribution, or the demand that would have arrived anyway. The model works the other way: channel spend is one regressor sitting alongside price, seasonality and competitor pressure, as the foundations lesson sets out, so it only sees what actually varies in the aggregate. If you spent within a narrow band in every one of the last 150 weeks, the model has nothing to learn from and will hand you a coefficient with an interval wide enough to drive a truck through.

The working split:

• Platform data for within-channel decisions: creative, audience, bid, placement. Nothing else has that granularity or that refresh rate.
• The model for cross-channel money: how much goes to search versus TV versus retail media next year.
• When the two disagree by more than about 2x on the same channel, treat it as an open question rather than averaging them. Splitting the difference is how you end up wrong by half in a direction nobody owns.

The hardest version of this arbitration is branded search and retail media, where the platform counts conversions that would have happened through an organic link at no cost. Those two lines are where model-led budgeting usually finds its first eight figures.

Calibrating the model with GEOGEOThe practice of making your brand and content visible and citable inside AI-generated answers from tools like ChatGPT, Gemini and Perplexity.View full definition → experiments

An unfalsified model is an opinion with regression output attached. Geo experiments are the cheapest way to falsify it: hold spend out of a matched set of markets for six to ten weeks, measure the sales gap, then check whether the model's coefficient for that channel predicted the gap you observed. If it did not, the model is wrong and you now know by how much and in which direction. Modern implementations feed the test result back in as a prior, so the next refresh is anchored to something you measured rather than something you fitted.

Two things a CMO has to protect in this process. First, the size of the perturbation: a 10% spend change in four markets typically cannot detect anything below a double-digit lift, so you either cut hard in the test cells or you accept an inconclusive result and stop pretending otherwise. Second, contamination: national TV, national price promotions and retailer decisions do not respect your test geography, and a competitor launching in your control markets can invert the reading. Read the test with the commercial calendar in front of you.

Budget for calibration as a standing cost, roughly one experiment per major channel per year. Rotate the channel under test. The channels that never get tested are exactly the ones your organisation has decided in advance to believe.

The model nobody in finance believes

This is the dominant failure mode, and it is organisational rather than statistical. The symptoms are recognisable well before the model dies:

• Contribution figures that cannot be reconciled to the P&L, because the model was built on gross media cost while finance books net of agency rebates.
• Coefficients that flip sign between refreshes, with no explanation offered beyond "the data updated".
• Point estimates presented with the intervals removed for readability. Finance finds out later, usually in the meeting where you ask for the reallocation.
• A vendor who will show a waterfall chart but will not hand over the elasticity table or the model specification.

The second-order consequence is worse than the wasted fees. Once a leadership team learns that a mix model can be re-specified until it agrees with the plan, the next vendor gets chosen for agreeableness, and the model becomes a ratification device for decisions already taken. Reversing that reputation takes two or three planning cycles.

The defence is boring and it works: reconcile the model's base sales to the actual P&L line before anyone sees a channel result, publish the interval next to every number, and never propose a reallocation smaller than the model's own uncertainty. If the interval on paid social ROIROIReturn on Investment: the ratio of net profit to the cost of an investment. A 300% ROI means each dollar invested returns $3.View full definition → runs from 1.1x to 2.4x, you cannot justify a 5% shift on that basis. You can justify commissioning a test.

Who owns the number

Independence is a structural decision, not a preference. The media agency that buys the inventory should not build the model that scores it. The same logic applies to the platforms: Meta publishes and maintains open-source tooling for both mix modelling and geo testing, which is genuinely useful, and Meta also sells the advertising those tools evaluate. Use the code, keep the specification and the test design under your control.

The same caveat runs the other way. Brand trackingBrand trackingRegular measurement of brand health metrics (awareness, image, preference, and purchase intent) over time, so shifts can be detected and linked to marketing activity.View full definition → vendors such as Kantar sell both the tracking series and the modelling services, so if you buy both from one supplier, ask who checks whose work. The tracking series itself is worth having in the model: a slow-moving brand equitybrand equityThe commercial value your brand adds beyond functional product attributes: the price premium, preference and loyalty it generates.View full definition → variable gives the regression something to attribute the persistent half of the effect to, instead of parking it in an unexplained baseline that grows every year.

Practically, name one internal owner (analytics or finance, not the media team), give them the vendor relationship, and put the model refresh on the same calendar as budget lock rather than four weeks after it.

Real-world cases

Case 1: eBay. Blake, Nosko and Tadelis, working with eBay's own economics team, ran a large-scale field experiment that switched off paid search advertising across a set of US markets and published the results in Econometrica in 2015. Branded keyword ads returned close to nothing, because the traffic reappeared through organic links at no cost, and the measurable returns on non-branded terms were concentrated among infrequent users. The platform reports had shown a healthy return the whole time. Every one of those clicks was real; the incremental sales largely were not. This is the sharpest available argument for why a model calibrated by experiment outranks a channel-reported number when the money moves.

Case 2: Meta. Meta open-sourced Robyn, a mix modelling code base that accepts lift-test results as calibration inputs, and GeoLift for designing the geo experiments that produce them. That a platform selling media publishes the machinery for measuring its own incrementalityincrementalityThe share of results (sales, conversions, revenue) that only happened because of a marketing action, not what would have occurred anyway.View full definition → tells you how contested the arbitration has become. Take the design pattern (experiment, then constrain the model, then allocate) and keep the referee outside the seller.

CMO action items

• Reconcile the model's base sales against the finance P&L before any channel result circulates, and agree with your CFO in writing which cost basis (gross or net) the model uses.
• Put one geo experiment per major channel per year into the plan, and commit in advance to acting on the result, including the case where it contradicts a channel you have publicly defended.
• Publish confidence intervals with every reallocation recommendation, and set a floor: no shift smaller than the interval.
• Deliberately vary spend in at least a few markets each year. A perfectly smooth plan starves the next model of the variation it needs to read anything.

Common mistakes that kill results

Mistake 1: Letting the agency that buys your media also build your model. The incentive is to show its channels performing. Buy the model from an independent measurement firm and give them clean first-party data directly.

Mistake 2: Rebuilding the model when the answer is unwelcome. If the specification changes because a result was politically difficult, you have destroyed the model's standing with finance permanently, and you will not get it back inside this planning cycle.

Mistake 3: Treating the model as an annual report rather than a decision instrument. If outputs land after budget lock, they become commentary. Line up the refresh so results arrive with enough time to argue about them.

Mistake 4: Optimising to the model's point estimate. Cutting a channel to the exact spend the curve prefers is precision the data does not support; move in steps you can measure, and re-test.