Why Your Marketing Mix Model Says Offline Doesn't Work (And Why It Might Be Wrong)

If you are responsible for marketing/media ROI, read previous posts first, and then think again: toys or tools for your teams? Both are okay, but you should know what you are buying and what for.

A company invests millions in a Marketing Mix Model. The results arrive in an impressive dashboard. Paid Search delivers an ROI of 6.8. Social Media is performing above expectations. Email is one of the highest-return channels (zero opex costs). Then comes the surprise.

Television and grandiose outdoor billboards, the company's largest marketing investments, appear to contribute almost nothing. You cannot hide media costs in the assets as you might do with production, even if some pressure makes you do so. The executive team's first reaction is predictable. "Why are we still spending money on TV? Let's put down that chocolate billboard from the supermarket parking lot wall?" Budgets shift. TV is reduced. Outdoor is moved indoor, into the office premises for a souvenir.

Performance marketing receives a larger share of investment. CRM email is expected to respond to all seasonality, opportunities, and innovation communications. For a quarter or two, everything looks fine. Then something subtle begins to happen: branded search volume starts declining, which means your organic channels appear less in attribution. Website traffic becomes more expensive. Search campaigns require higher bids to generate the same number of conversions. Remarketing audiences shrink. Customer acquisition costs increase. Suddenly, every "high-performing" digital channel is working harder to produce the same results. Nothing is broken. Demand simply isn't being created anymore.

The offline campaign wasn't failing. It was feeding the entire marketing ecosystem.

The Marketing Mix Model simply couldn't see it.

Marketing Is an Ecosystem, Not a Competition of Channels or Managers

One of the biggest mistakes organizations make is treating marketing channels like competitors fighting over sales. They're not. They're teammates.

Think about how consumers actually buy. Very few people watch a television commercial and immediately purchase. Instead, the commercial plants a seed. A few days later they notice the brand again on YouTube. Later they see a friend mention it on LinkedIn. A week later they search Google. Eventually they click a paid search ad and buy. When analysts look only at the final transaction, Google Search appears to deserve all the credit. But Search didn't create the demand. It harvested demand that another channel had already created. This distinction is critical. Marketing channels don't simply generate conversions. They influence one another.

The Invisible Value of Awareness

Performance marketers often prefer channels that generate measurable clicks. It's understandable. Clicks are easy to measure. Conversions are easy to attribute. Awareness is much harder.

Imagine asking someone why they chose Coca-Cola instead of an unfamiliar supermarket brand. Few people can point to one advertisement. Instead they say, "I've always known the brand." That familiarity didn't appear overnight. It accumulated through years of advertising, sponsorships, product placement, retail visibility, conversations, and repeated exposure. Brand advertising rarely creates immediate transactions. Instead, it lowers the mental effort required to choose later. That value is difficult for any statistical model to isolate.

Correlation Is Easy. Causation Is Hard.

Marketing Mix Models work by looking for relationships in historical data. Suppose TV spend and sales often increase together. Does television cause higher sales? Maybe. But perhaps TV campaigns always coincide with holiday promotions. Or new product launches. Or increased retail distribution. Or price discounts. Or seasonal demand. Without accounting for these factors, the model may assign too much or too little credit to television. This is known as omitted variable bias. The model is forced to explain sales using only the variables it can see. Anything missing gets absorbed into the variables that remain. Sometimes TV gets too much credit. Sometimes it gets almost none. Neither necessarily reflects reality.

When Channels Move Together, Statistics Get Confused

One of the least discussed challenges in Marketing Mix Modeling is multicollinearity.

It sounds technical, but the idea is simple. Imagine your company launches campaigns this way: TV increases. YouTube increases. Paid Search increases. Display increases. PR activity increases. Every campaign starts at roughly the same time. From a statistical perspective, all those variables move together. Now ask any model: "Which one caused the increase in sales?" It can't confidently answer. It's like trying to determine which musician deserves credit for an orchestra's performance. The violin, piano, and cello all played simultaneously. Removing one changes the music, but measuring each individual contribution becomes extremely difficult.

When marketing channels are highly correlated, even sophisticated models struggle to separate their individual effects. The output may appear mathematically precise. The underlying uncertainty is often much larger than the report suggests.

The Creative Problem No Model Can Solve

Here's another assumption hiding inside many MMM implementations: every euro spent on TV is treated as roughly equivalent. Anyone who has worked in marketing knows this isn't true. A brilliant commercial and a forgettable commercial cost the same to air. The model sees identical media spend. It tracks exact online clicks a few minutes or hours from airing. Very vague timing, usually depends on the shopping habits of your managers. Consumers experience completely different advertising. Creative quality, messaging, emotional resonance, and cultural relevance rarely exist in marketing datasets. Yet they often explain why one campaign transforms a business while another quietly disappears. The model measures exposure. Customers respond to creativity. Those are not the same thing.

The Long Memory of Advertising

One of the most difficult aspects of brand advertising is that its effects unfold slowly. A search ad may generate a conversion within minutes. A video campaign may influence customer perceptions for months. Marketing Mix Models try to account for this using adstock — the idea that advertising has a carryover effect that fades over time. But here's the challenge: How long does a TV commercial remain in someone's memory? Two weeks? Six weeks? Three months? How many OTS is pleasant and when is it annoying? Is it okay to be annoying but memorable? Or is it just throwing water in the river? Once the message is floating in harmony with the consumer, there is no need for flooding. Crazy OTS means mostly that TV airing was on discount, so someone purchased it with the full budget amount — it would be strange to offer the budget left to another channel. But you can't save time for later like a big pack of cookies. Keep that in mind and better ask for share of voice. OTS means you went to the same party x times, but you don't know if anyone noticed. Share of voice shows if there was a possibility that someone actually noticed you out of the crowd. There is no universal answer. Every model must assume a decay rate. Different assumptions produce different results. This is why two perfectly competent analysts can build two different MMMs from the same data and reach different conclusions. Neither is necessarily wrong. They simply made different assumptions about how consumers remember advertising.

The Danger of Optimizing for What You Can Measure

Perhaps the greatest risk isn't that MMM underestimates TV. It's that organizations gradually shift investment toward channels that are easiest to measure. Search. Email. Retargeting. Affiliate marketing. These channels often look exceptional because they intercept consumers who are already close to purchasing. Over time, companies become increasingly efficient at harvesting demand. Meanwhile, fewer investments are made in creating new demand. It's similar to a farmer harvesting crops while gradually planting fewer seeds. For a while, everything looks productive. Eventually, the harvest becomes smaller.

The problem wasn't harvesting. It was forgetting to plant.

So, Should You Ignore Marketing Mix Models?

Absolutely not. Marketing Mix Modeling remains one of the most valuable tools available for understanding long-term marketing performance — especially in a privacy-first world where user-level tracking is becoming less reliable. But it should never be treated as a verdict. Think of it as evidence. Sometimes very strong evidence. But still only one piece of evidence. The strongest organizations combine Marketing Mix Models with incrementality experiments, geo-lift tests, brand tracking studies, customer surveys, CRM analysis, and digital attribution. When multiple methods point in the same direction, confidence increases. When they disagree, curiosity should increase, not certainty. Chaos means you need more research, not going back to last year's copy-paste.

The Best Marketing Leaders Ask Better Questions

Instead of asking, "Does TV work?" ask:

Does TV increase branded search?

Does TV improve the performance of paid search?

Does TV shorten the customer journey?

Does TV increase direct traffic?

Does TV improve conversion rates across other channels?

What happens when TV is removed — not immediately, but six months later?

These questions acknowledge an important truth. Marketing channels do not operate in isolation. They interact. They reinforce one another. Sometimes their greatest contribution is making every other channel more effective. The most dangerous number in marketing is not an inaccurate ROI. It's an ROI presented with unwarranted certainty.

Consumers don't experience marketing as separate channels. They experience a brand. Every impression, conversation, recommendation, search, store visit, and advertisement contributes to that experience. Marketing Mix Models attempt to simplify this complexity into a set of coefficients. Sometimes they do an excellent job. Sometimes they don't. The responsibility of the analyst isn't to defend the model. It's to understand where the model is likely to be wrong. Because in marketing, asking "What if the model is missing something?" often leads to better decisions than believing the model has found the final answer.