How a Single Harrah’s Property Cut More Than $30 Million a Year Without Losing Performance
In 2012, working in partnership with Polaris Research, we used marketing mix modeling to study and optimize the promotional campaigns at a single Harrah’s property. The property cut more than $30 million a year from its marketing spend, and performance held.
That result is more than a decade old. I’m telling it now because the problem it solved is back, and bigger. Clicks no longer show where a sale started. The measurement that works without them is the same kind that found that $30 million. It’s also the backbone of chapter 7 of our guide, on measuring what grows revenue when clicks disappear.
What happened at the Harrah’s property?
The work focused on one property’s promotional campaigns and one question: which promotions were earning their keep, and which only looked like they were?
Every big marketing budget carries programs like these. Each one has a champion, a history and a report that says it worked. Reports like that are easy to believe and hard to check, because every program takes credit for customers who might have come anyway.
Mix modeling doesn’t ask each promotion to grade itself. It looks at spend, results and outside factors over time, all at once, and estimates what each piece of spend added. Then you can move money on evidence instead of habit.
At this property, that work cut more than $30 million a year from marketing spend, with performance maintained. That’s the whole claim, and it’s the one I can stand behind.
What is marketing mix modeling, in plain English?
Marketing mix modeling (MMM) is a statistical model that estimates how much each channel or campaign adds to sales, using totals over time instead of tracking individual people.
That last point is the one that matters in 2026. Attribution tools follow a person from click to click. When the click goes away, so does their view. A mix model never depended on it.
Why does a 2012 result matter now?
Because the click is going away, and it’s taking click-based measurement with it.
Pew Research Center tracked the browsing of 900 U.S. adults in March 2025. When Google showed an AI summary, people clicked a regular search result on 8% of visits. Without one, they clicked on 15% (Pew Research Center). Ahrefs found that an AI Overview goes with a 58% lower average clickthrough rate for the top-ranking page (Ahrefs).
Every one of those missing clicks is a gap in someone’s attribution report. The sale may still happen. Your dashboard just can’t see where it started. So the question a CFO asks, “what can we cut without hurting revenue?”, gets harder to answer with the tools most teams use.
Companies know this. In a Harvard Business Review survey sponsored by Google, 73% of respondents said their organizations are under growing pressure to show the effects of their media channels. Yet only 22% said they’re “very effective” at using MMM to act on what it shows (EMARKETER).
Which measurement answers which question?
No single method does everything. The mistake is using the one you have for a question it can’t answer.
| Method | The question it answers | What it needs | Where it goes blind |
|---|---|---|---|
| Click-based attribution | Which tracked touch came before a sale? | User-level clicks and cookies | AI answers, offline sales, anything without a click |
| Marketing mix modeling | What did each kind of spend add, and what if we move money? | Totals over time for spend, sales and outside factors | Fast changes inside one campaign; very new channels with little history |
| Incrementality test | Did this specific spend cause extra sales? | A test group and a holdout group | Only one question at a time; takes planning |
The strongest setups use the model to find candidates for cuts and a test to confirm the big ones. Google’s Meridian, for example, can take incrementality experiment results as inputs (“priors”) to the model (Google).
Why do so few companies act on what the model says?
Because a model doesn’t cut anything. People do, and most organizations aren’t set up to let a model change their minds.
In that same HBR survey, the reasons for the gap between insight and action were human: 46% cited slow processes, 45% a lack of expertise and 41% siloed teams. Leaders were far more likely than laggards to connect model outputs to real marketing decisions: 55% versus 20% (EMARKETER).
That matches what I’ve seen over 35 years in marketing, since 1991. The math is rarely the hard part. The hard part is the meeting where a model says a beloved program isn’t earning its keep.
What should a CMO take from this in the AI era?
Three things, and none of them is “buy a model.”
Decide before you model. Write down the budget decision the model has to inform and the rule you’ll follow when it answers. “If a promotion adds less than it costs for two quarters, we cut it” is a rule. “Let’s see what the data says” isn’t.
Measure the parts AI made invisible. Your mix should include what clicks can’t see now: being named in AI answers, earned media, events and sales conversations. If it isn’t in the model, the model will hand its credit to whatever is.
Let AI speed up the work, not make the call. The tools are moving fast. Google’s Meridian page now lists an agent toolkit that helps analysts run the model and fix errors (Google for Developers). That makes models cheaper and faster to build. It doesn’t decide which program to cut, or take the heat for it.
What would this look like for your budget?
Start smaller than you think. Pick one decision you have to make in the next two quarters, such as what to cut, what to protect and what to test. Then ask whether your current measurement can answer it without clicks. If it can’t, that’s the gap to close first.
If you want a second set of eyes on that, ask about the measurement and AI-visibility audit. It’s a fixed-fee review of which spend grows revenue and where AI answers name or miss your brand, with written decision rules at the end.
A promotion that grades itself always passes. Make it take the test.
