Your Best-Looking Ad Channel May Be Doing the Least: A CMO’s Guide to Marketing Mix Modeling

In a two-year simulation, last-click attribution said branded search returned $13.26 per dollar. The true number was $1.20. What the evidence says, what a marketing mix model measures that attribution can’t, and how to act on it.

Every CMO eventually gets the same question from the CFO: if we cut this channel in half, what happens to sales? Most marketing dashboards can’t answer it. They report which ads collected the clicks and the credit. That is a different question, and the gap between the two answers is often where budget goes to waste.

Marketing mix modeling was built for the CFO’s question, and it is back. Apple’s App Tracking Transparency arrived with iOS 14.5 in April 2021, privacy rules cut into tracking, and a decade of experiments showed how often click paths credit ads with sales they didn’t cause. In a 2025 EMARKETER and TransUnion survey, 46.9% of US marketers said they plan to invest in marketing mix modeling over the next year. I’ve spent more than twenty years in marketing modeling and measurement, and this is the condensed version of a longer guide I wrote for marketing leaders. The full CMO playbook is on Substack.

Attribution measures credit, not cause

Attribution answers one question: which tracked touchpoint came before this sale? It can’t see TV, audio or billboards, so they get zero credit by construction. And it has no counterfactual. A customer who was going to buy anyway, searched your brand name and clicked the ad at the top of the page counts as a paid search conversion.

That second gap has been measured. When eBay paused brand-keyword ads on Yahoo and MSN in a controlled experiment, organic results recovered 99.5% of the clicks. For non-brand keywords, the researchers estimated a short-term return of −63%, because most of the spend reached frequent buyers who would have purchased anyway.

The evidence keeps pointing the same way

At Facebook, researchers compared standard measurement methods against 15 randomized experiments. In one checkout study the true lift was 77%. A simple comparison of people who saw the ad with people who didn’t put it at 416%, about 5.4 times too high.

Even clean experiments struggle. In 25 large retail advertising experiments, Lewis and Rao found the typical confidence interval on return on investment spanned more than 100 percentage points, because ad effects are small next to the normal swings in sales. If a randomized test strains to see the effect, a click path with no control group has no chance.

A study of TV advertising for 288 consumer brands found a median elasticity of 0.01: doubling TV exposure raised sales about 1%, and weekly return on investment was negative for more than 80% of products. Uber’s former head of performance marketing has described turning off roughly $100 million a year in ads with no meaningful drop in app installs. P&G cut $200 million in digital spend in 2017 and said its reach rose 10%. Airbnb cut marketing 58% in 2020 and kept traffic near 95% of the prior year.

Those company stories have caveats. Uber’s cut removed fraud, P&G reported reach rather than sales, and Airbnb’s cut came during a pandemic. None of them prove advertising doesn’t work. They show something narrower: the channels that look best in attribution reports are often the ones doing the least, because they sit closest to customers already on their way.

Same data, two answers

To see the gap in numbers, I simulated two years of weekly sales for a brand running TV, paid social and branded search, with every channel’s true return known. Then I measured it two ways.

• TV: true return 4.77 per dollar. Last-click said 0.00. The mix model said 4.83.
• Paid social: true 3.57. Last-click said 2.69. The mix model said 4.00.
• Branded search: true 1.20. Last-click said 13.26. The mix model said 1.36.

Last-click got the order exactly backwards. 32% of the sales TV created were credited to branded search, and only 9% of the credit search received was sales its ads caused. A budget built on that report would cut the channel doing the most work and fund the one doing the least.

What a mix model actually does

A mix model explains weekly sales as a baseline plus each channel’s contribution, holding season, price and promotions fixed. It needs no user tracking. Two ideas sit at its core.

Because it works from aggregate totals, a mix model also tells you how much of your sales would happen with no advertising at all. In the simulation that baseline was about 72% of sales, which is typical of an established brand and is the single number attribution hides most completely.

The first is carryover. Advertising keeps working after the week it runs. If 60% of the effect carries into each following week, a single $100,000 week behaves like $250,000 of spend spread over the weeks after it.

The second is diminishing returns. Each extra dollar reaches fewer new people. In the simulation’s TV curve, $150,000 of effective weekly spend added $300,000 in sales, and doubling it to $300,000 added only $100,000 more. That is why the return on the next dollar, not the average dollar, should drive budget decisions.

Where mix models go wrong

A mix model is a regression on a small dataset, and it can be confidently wrong. When I reran the simulation 20 times, paid social’s estimated return ranged from 2.59 to 4.22. When the branded search budget followed demand, as many teams set it, the model put search at 2.03 against a true 1.20.

The fix is experiments. Turn a channel off in a set of matched markets for a few weeks, measure what happens, and feed the result back into the model. Pausing a $3 million branded search budget in 20% of markets for six weeks pauses about $69,000 of spend, small next to a decision about $3 million a year.

Three things to do this quarter

Ask a different question in your next budget meeting. Not which channel drove the most conversions, but what sales would have been if you had spent nothing there. If no one can answer for any channel, that is the gap to close.

Test the channel attribution flatters most. For most brands that is branded search or retargeting. A regional holdout for a few weeks costs little and gives you a causal number no dashboard can argue with.

Demand ranges, not single numbers. Whether the answer comes from a vendor, an agency or your own team, a channel return without an uncertainty range is a guess presented as a fact. A return of 2.0 with a range of 1.6 to 2.4 can move money. A return of 2.0 with a range of 0.3 to 4.5 should go to the top of your test list.

Keep attribution for what it does well: fast feedback on creative, keywords and audiences inside a channel. Just don’t let it decide how much each channel gets.

What the full CMO playbook covers

The Substack guide goes further, with the tools a marketing leader needs to act:

• A readiness scorecard, including the data-points rule that tells you whether your history can support a model at all
• Build, buy or hire, with published cost ranges for in-house teams, platforms and agencies, and a break-even formula. For example, a $150,000 program on $10 million of media pays for itself if just 3% of spend is misallocated.
• Ten questions to ask any vendor, agency or internal team before you commit
• How to read a model’s output, with seven red flags of a bad one
• How to run the experiment program: what to test first, what it costs, how often
• How to bring the CFO along and how fast to move money
• A week-by-week 90-day plan

It’s written for CMOs, marketing finance partners and measurement leads who need answers they can defend in a budget meeting.

Read the full CMO guide on Substack → HERE

About the Author

Brian James Curry is a Kansas City AI researcher, applied scientist and data scientist with more than twenty years in marketing modeling and measurement. His work today spans marketing mix modeling, causal inference, machine learning and healthcare intelligence. He is the founder of Vector1 Labs, an independent applied AI research lab. His marketing measurement work covers marketing mix modeling, incrementality testing and budget allocation, and his Causal Care series applies causal machine learning to health system decisions: which programs work, for whom, and what they are worth. He also founded KC AI Lab, which began in 2016 as the Kansas City Machine Learning Group and has grown to more than 1,500 members.

His career includes leading analytics, data science and machine learning at Hallmark, Garmin, Koch Industries, Tractor Supply Company and Vail Resorts. His open-source research tools on GitHub include Papilon (marketing mix modeling and complex systems), PyCausalSim (causal discovery through simulation) and tincast (forecasting for AI-critical commodities), along with MeaningFlow and Daedalus. A Kansas State University alum from a small Kansas farming town, he writes about marketing mix modeling, causal inference, agentic AI, healthcare AI and the economics of AI on Medium and Substack. Contact: brian@vector1.ai.