Your Marketing Model Is Only as Good as Your Data (and Yours Is Probably Wrong)

The Meeting That Changed Everything

It started as a routine data review. A national franchise retailer with over 100 stores had been running a marketing mix model for several months. The results were interesting. The channel recommendations were useful. But something felt off.

Then their media agency’s CEO joined the call for the first time.

Within ten minutes, he spotted the problem that had been hiding in plain sight: the data going into the model was wrong.

Not slightly wrong. Search spend was showing $100,000-$200,000 when the actual spend was $200,000-$300,000. That’s a 50-100% undercount on one of their most important channels.

How the Data Got Broken

The root cause was simple but surprisingly common. The marketing measurement system was pulling data from two different sources that told different stories.

Digital channels came through an API (Supermetrics) that pulled directly from publisher platforms like Meta, Google, and Yahoo. This sounds accurate, and it often is, but only if every account is connected, every campaign is mapped correctly, and the publisher’s numbers match what the agency actually delivered.

Offline channels (TV, radio, out-of-home) came from the media agency’s planning documents. Not actuals. Plans. And as anyone who’s worked in media knows, plans and actuals can differ significantly.

The agency CEO put it bluntly: “Best practice in the modelling that we have worked across is that we’re looking at actuals, not planned, because they can differ quite significantly.”

Three Specific Problems They Found

1. The Yahoo Black Hole

All spend through Yahoo’s DSP was being categorised as “Yahoo” in the model. But Yahoo was actually buying multiple different media types: BVOD (broadcast video on demand), digital audio, and display.

Lumping them together meant the model couldn’t tell which format was actually working. The Head of Marketing said it plainly: “The problem with that is it doesn’t help us to know which levers to turn on and off if Yahoo is different medias.”

The fix: split by media type (BVOD, audio, display) instead of by source (Yahoo). Simple, but nobody had caught it because the data was being ingested automatically.

2. The Missing YouTube Channel

An entire YouTube channel’s spend was completely absent from the model. It was being bought through DV360 (a different Google platform), and the API connection to Supermetrics only covered the main Google Ads account.

Tens of thousands of dollars in video spend were invisible to the measurement system. Not underreported. Invisible.

3. The Meta Objective Gap

Meta spend was being reported as one lump number. But the agency was running three distinct campaign types: awareness, consideration, and conversion. Each has fundamentally different objectives and expected returns.

Measuring them as a single channel is like measuring “transport” as one category when you’re actually running taxis, buses, and freight. The economics are completely different.

The Fix: Let the Agency Own the Data

The solution wasn’t more technology. It was a process change.

Instead of the measurement team pulling data from APIs and interpreting media plans, the agency would provide actuals directly via a standard template. The agency CEO explained why: “We’re the ones closest to the data and what’s actually being delivered.”

The new workflow had three principles:

Actuals only. No more media plans as a data source. What was actually spent and delivered, not what was planned to be spent.

Exclude noise. All spend reported ex-GST, excluding agency fees and production costs. These don’t affect media performance, and including them inflates the apparent spend without changing the signal.

Agency validation before modelling. A new pre-review step where the agency checks the data before it goes into the model. If there are discrepancies between what the API shows and what was actually delivered, the agency catches it first.

Why This Matters More Than You Think

Marketing mix models are only as good as their inputs. A model built on inaccurate data will confidently give you wrong answers. It will tell you to invest more in channels that look efficient (because their spend is undercounted) and less in channels that look expensive (because their spend is overcounted).

In this case, search was being underreported by up to 50%. If the model used that data to recommend budget allocation, it would have overestimated search ROI and recommended pouring more money into a channel that was already spending more than the model knew.

The scary part: the model still looked reasonable. It still had good accuracy metrics. The outputs were plausible. Without the agency CEO’s institutional knowledge, nobody would have questioned it.

The Broader Lesson for Every Brand

This problem isn’t unique to one retailer. It exists anywhere that:

• Multiple data sources feed into a measurement system
• API data and manually reported data need to be reconciled
• Media is bought through platforms (Yahoo, DV360) that aggregate multiple media types
• Campaign structures don’t map cleanly to measurement categories

If you’re running any kind of marketing measurement, ask yourself three questions:

1. Are you measuring actuals or plans? If your offline data comes from media plans rather than delivery reports, your numbers are probably wrong.

2. Does your measurement category match your decision category? If you make decisions by media type (video, audio, display) but your model groups by platform (Yahoo, Google), you can’t action the results.

3. When was the last time your agency validated the data going into your model? If the answer is “never,” that’s your next meeting to schedule.

The Takeaway

The most sophisticated model in the world can’t overcome bad inputs. Before you optimise your channel mix, optimise your data pipeline. And bring your agency into the room. They’re closer to the delivery data than anyone else, and they’ll spot problems that no API or dashboard will surface.