Real-world application of marketing mix modeling

In October 2019, Adidas' global media director Simon Peel told an industry audience that the company had over-invested in digital performance marketing and under-invested in brand building. The split he put on screen: roughly 77% of media money going to performance, 23% to brand, against an econometric read that put something like 65% of sales on the brand side of that line. This lesson stays inside that one review: the data that went into it, why the reporting Adidas had been using pointed the opposite way for years, and what it costs a company to act on a finding it does not like.

The starting position was deliberate, not sloppy. In 2016, chief executive Kasper Rorsted had told CNBC that consumer engagement happens mostly on mobile, and Adidas leaned into digital hard enough to skip TV for the 2018 World Cup. Every weekly performance report the media teams saw endorsed that choice, because every weekly report was built out of click paths.

First, the data build, which is where a review like this either earns credibility or loses it. Adidas sells in more than 100 markets, and in 2019 more than half of roughly €23.6bn in revenue came through wholesale, meaning the sale happens in someone else's store with a lag and often without clean sell-out data. So the model cannot be fed group revenue. It needs weekly series per market: media spend by channel, price and discount depth, distribution and product launches, competitor pressure, plus the football calendar, which moves demand independently of media. Markets where sell-out data is thin get modelled separately or dropped, and the honest version of the exercise says which ones were dropped.

Second, the reason the old numbers lied. Adidas' digital reporting credited the final click, and much of the media was optimised toward conversion on adidas.com, while own ecommerce ran at roughly a tenth of group revenue. Paid search, affiliates and retargetingretargetingShowing ads to users who have previously visited your site or interacted with your brand, to bring them back and drive conversion.View full definition → were being paid for demand that brand activity and the retail network had already created, then re-purchased at the checkout door. The same money looked like a 10x return in the platform view and something close to nothing in the econometric view, and the gap was not a modelling artefact. It was the difference between measuring who was present at the purchase and measuring who caused it.

Third, the identification problem that makes this finding hard to produce at all. Adidas ramps brand and performance spend together around launches and tournaments, so the two series move in step and the model struggles to separate them. If your brand and performance budgets have never diverged, no amount of adstock fitting will tell you which one did the work. What rescues it is variation: markets that launched a campaign three weeks apart, a country that went dark on TV for budget reasons, a staggered rollout. Practitioners who skip this step and report a clean brand coefficient from perfectly collinear data are reporting their prior, not a result.

Fourth, the window problem, and this is the part that flatters performance even after you accept the persistent decay profile the foundations lesson describes. A model fitted on two or three years of weekly data can only see media whose payback lands inside that window. Brand investment that shows up as pricing power or lower search costs in year four sits in the base, indistinguishable from heritage. So even a review that lifts brand from 23% of spend to a majority of measured effect is probably still understating it, which is the uncomfortable direction of the error. Les Binet and Peter Field's long-and-short work put the practical balance nearer 60:40 brand to activation for exactly this reason.

Acting on the finding is worse than producing it. If you move money out of paid search and affiliates in Q1, the revenue those channels were harvesting drops immediately, because they were closing demand that still needed closing. The brand payback arrives over quarters. So the P&L gets worse before it gets better, and the person who authorised the shift owns that gap in front of a board that read the old dashboard. There is a margin dimension too: performance channels often carry discount codes, so part of what the model was crediting to media was price. Rebalancing toward brand can hold revenue flat while gross margingross marginGross margin is the share of revenue left after subtracting the direct cost of producing goods or services, expressed as a percentage of revenue.View full definition → improves, which never appears in a media report at all.

The second-order consequence Adidas ran into is organisational. A digital-heavy split builds a digital-heavy team, digital-heavy agency contracts and weekly rituals around metrics the review has just demoted. The model output arrives as an attack on several people's last three years of work, which is why Peel framed it as a measurement failure rather than a competence failure.

CMO Action Items:

• Before commissioning anything, write down what share of your sales your own ecommerce actually carries. If it is 10% and your media is optimised to on-site conversion, you already know the direction of your bias, and the model is there to size it, not discover it.
• Insist on market-level models, not one global model. Adidas' review only had usable signal because different markets spent differently at different times. A single blended series hides the variation the estimate depends on.
• Put price and promotion in the model as separate variables before you argue about channels. Otherwise discount-carrying channels absorb credit that belongs to margin you gave away.
• Sequence the rebalance across two or three quarters and tell finance the revenue dip is expected, with a size attached, before it happens. An unforecast dip in month two is how a correct decision gets reversed.

Common Mistakes That Kill Results:

• Treating a rising platform-reported return as evidence against the model. The two numbers measure different things, and a channel sitting at the point of purchase will always look strong in a last-click frame. Reconciling them is a leadership question the playbook lesson takes up.
• Letting the data window set brand's value permanently. Two years of history structurally under-credits slow-burn investment, so a model that says brand is worth 40% is a floor, not a verdict. Rerun with a longer series when you have one.
• Accepting a model built on collinear spend. If brand and performance budgets have always moved together in your business, deliberately break the pattern in a few markets for a quarter so the next model has something to read.
• Filing the review as a one-off. Media costs, price architecture and distribution all shift; a split that was right in 2019 is a guess by 2022. Recalibrate annually and keep the same specification so the year-on-year comparison means something.