Why Marketing Mix Models Beat Attribution for Long Sales Cycles

Lucas: Last month I was looking at attribution data for a B2B SaaS client with a six-month sales cycle, and the last-touch model said their biggest driver was a retargeting ad that ran in week twenty-four. But when I looked at the raw CRM data, that prospect had consumed eighteen pieces of content before they ever saw the ad - whitepapers, webinars, case studies - and none of those got any attribution credit. Luna: That is the core problem with attribution for long cycles. It gives all the credit to the last click, or maybe a weighted fraction if you use multi-touch, but it still misses the early-stage influence entirely.

Lucas: Exactly. And that's why I want to talk about marketing mix models today. Not as a complete replacement for attribution, but as a tool that handles long time horizons and aggregate effects way better. Let's use a concrete example.

Say you sell a fifty-thousand-dollar annual contract, and the average buyer interacts with your brand twenty-two times before signing. An attribution model, even a sophisticated one, can only see the touchpoints it's tracking - clicks, form fills, email opens. It can't see the brand awareness from that podcast ad they heard eight months ago, or the industry report they read that mentioned your company. Luna: Marketing mix modeling does capture that, though, right?

Because it works at the aggregate level, looking at total spend across channels over time and correlating it with sales outcomes. Lucas: Right. Instead of assigning credit to individual touches, an MMM estimates how much each channel contributes to overall revenue, controlling for things like seasonality, competitive activity, and economic conditions. A 2024 study from the Institute for Marketing Science looked at fifty-two companies with sales cycles longer than ninety days and found that MMMs outperformed attribution models by about thirty percent in forecasting accuracy.

Luna: Thirty percent is significant. But I imagine there's a trade-off. MMMs require a lot more data and statistical expertise. They're not something you can spin up in a week.

Lucas: No, they're not. You typically need at least two years of weekly or monthly data on spend, sales, and external factors. And you need someone who can build and maintain a Bayesian or econometric model. That's a bigger investment than plugging in an attribution platform.

But for companies with long cycles and large marketing budgets, the insight can be worth it. Luna: And if these marketing conversations have sparked something you've actually used in your own work, listener support is what keeps this show ad-free and independent. You can find us at buy me a coffee dot com slash fexingo. Every little bit helps.

Lucas: Absolutely. And speaking of practical use, let's talk about how attribution and MMMs can actually complement each other. The best teams I've seen run both in parallel. They use attribution for granular, tactical optimization - like which ad creative is working this week - and MMM for strategic budget allocation across quarters.

Luna: That makes sense. Attribution gives you speed and granularity, but MMM gives you the big picture. But I have a question: doesn't MMM have its own blind spots? For example, if you're launching a new product, you don't have historical data.

Lucas: Great point. MMMs are backward-looking by nature. They need history. So for new products or new channels, you can't rely on the model alone.

That's where incrementality experiments or geo tests come in. You run a controlled test to measure the causal impact of the new channel, and then feed that result into the MMM as a prior. Luna: So it's not one or the other. It's a toolkit.

And the size of the company and the length of the sales cycle should dictate which tool gets more weight. Lucas: Exactly. If you're a direct to consumer brand with a two-day purchase cycle, last-touch attribution might actually be fine because the consideration window is short. But if you're selling enterprise software or medical devices or capital equipment, you need a model that can see across months.

Luna: Let's get into the mechanics a bit. How does an MMM actually separate the effect of a channel from, say, seasonality or competitor activity? Because those things are correlated. Lucas: Good question.

The model includes variables like month-of-year, GDP growth, competitor ad spend if you have it, even weather or holidays. It uses Bayesian methods to estimate the marginal contribution of each channel while holding everything else constant. For example, if you spend more on trade shows in Q3 and sales go up in Q4, the model can estimate how much of that lift is due to the trade shows versus the normal seasonal uptick. Luna: But how do you know the model is right?

I've seen teams overfit MMMs, adding so many variables that the model fits the past perfectly but predicts the future terribly. Lucas: That's a real risk. The best practice is to hold out the most recent quarter of data during training, then test the model's predictions against actuals. If the forecast is off by more than ten percent, you probably overfit.

Also, keep the model simple. Often a model with five to seven channels plus three or four external factors is more robust than one with twenty variables. Luna: And what about the 'halo effect' that attribution misses? For example, a display ad might not drive a click, but it increases branded search volume.

An MMM can capture that as a cross-channel lift. Lucas: Yes. That's one of the biggest advantages. In attribution, if a channel doesn't get clicked, it gets zero credit.

But in reality, that display ad built awareness that later converted through a search ad. An MMM can estimate the halo by looking at how changes in display spend correlate with changes in search conversion rates. One client I worked with found that their display ads had a halo effect of thirty percent on search - meaning for every dollar spent on display, search got an extra thirty cents worth of conversions. Luna: That's huge.

So if you cut display based on attribution alone, you'd lose that halo and not even realize it. Lucas: Exactly. And that's the danger. Attribution can lead you to underinvest in top of funnel channels because they don't show a direct return in the short window the model tracks.

MMMs, because they look at aggregate relationships, are better at capturing those delayed and indirect effects. Luna: Let's talk about the cost. A proper MMM, built by a consultant or an internal data science team, can run anywhere from fifty thousand to a couple hundred thousand dollars. That's not nothing.

For a mid-sized company, is it worth it? Lucas: It depends on your marketing budget. A rule of thumb I've heard: if you spend more than five million dollars annually on marketing, an MMM will probably pay for itself by reallocating just five to ten percent of spend more efficiently. For a ten million dollar budget, that's half a million to a million in improved ROI.

So the math works. Luna: But what about companies that can't afford a full MMM? Are there lighter-weight alternatives? Lucas: Yes.

You can run a simpler time-series regression using just your own spend data and a few external factors. That's not a full MMM, but it can give you directional insights. Or you can use a tool like Facebook's open-source Robyn, which is a semi-automated MMM package. It's free, but you still need someone who can work with R or Python to set it up.

Luna: I've heard that some companies use attribution as their 'source of truth' for optimization even when they have an MMM. Is that a mistake? Lucas: Not necessarily, if you understand the limitations. Attribution is great for real-time optimization - like pausing a campaign that's underperforming this week.

But for quarterly budget planning, you should use the MMM. The key is to reconcile the two. If your attribution says display has a 0.5 ROAS but your MMM says it has a 2.

0 ROAS including halo effects, you need to decide which number to trust for planning. Luna: And how do you reconcile? I've seen teams average them or just pick one, which seems arbitrary. Lucas: The better approach is to dig into the discrepancy.

Why are they different? Maybe the attribution model has a short lookback window, or it misses offline conversions, or the MMM is misattributing seasonality to a channel. Use the gap as a diagnostic tool. If they diverge wildly, it often points to a data quality issue or a modeling assumption that needs revisiting.

Luna: I want to ask about the future. With the deprecation of third-party cookies, attribution is getting harder. Does that make MMMs more attractive? Lucas: Absolutely.

Without cross-site tracking, attribution models lose the ability to follow users across the web. MMMs don't rely on user-level data at all - they use aggregate spend and sales. So they're immune to cookie deprecation. That's a huge advantage going forward.

Luna: But MMMs still have their own challenges. Data silos, for one. If your sales data is in Salesforce and your ad spend data is in a dozen different platforms, getting it all into one model is a big data engineering task. Lucas: No question.

That's often the hardest part. But once you have the pipeline, the model can run automatically on a monthly or quarterly basis. And the insights compound over time as you accumulate more data. Luna: So if someone listening is thinking about moving toward MMMs, what's the first step?

Lucas: First, audit your current attribution. Understand where it's likely misleading you - especially for channels that have long lag effects or halo effects. Then, gather two years of monthly data: spend by channel, revenue, and any external factors you can get. Even if you don't build a model yet, just seeing the data in a spreadsheet can reveal patterns.

Then consider a pilot MMM for your longest-cycle product line. Start simple, validate the outputs against your intuition, and iterate. Luna: Good advice. And I'd add: don't abandon attribution entirely.

Use it for tactical decisions, but use MMM for strategy. The two together give you a more complete picture than either alone. Lucas: Exactly. And that's the takeaway: no single model gives you the truth.

But by understanding the strengths and weaknesses of each, you can make smarter decisions about where to invest your marketing dollars.