Latest / The CMO Podcast with Fexingo: Marketing Leadership, Budgets, and Executive Strategy / How CMOs Use Causal AI to Find Hidden Growth Levers
Transcript
- Lucas: So Luna, I keep seeing this phrase 'causal AI' pop up in martech decks, and honestly, I'm a little suspicious. Is it just another buzzword, or is there something real here? Luna: I had the same reaction. But then I talked to a data science lead at a big athletic apparel brand, and she showed me a story that made me a believer. Lucas: Okay, I need to hear this. What happened? Luna: So they had this classic problem. Their loyalty emails were getting hammered for being 'low funnel' and 'boring,' while their social campaigns were all about brand love and engagement. The CMO was about to cut email budget and pour more into social. Lucas: And the dashboards backed that up? Luna: Exactly. The social campaigns had massive reach, tons of likes, great click-through rates. The emails had maybe a two percent open rate and almost no clicks. On paper, it was a no-brainer. Lucas: But the paper was lying. Luna: Completely. They ran a causal analysis — basically a bunch of holdout tests and some fancy modeling — and it turned out the loyalty emails were driving something like thirty percent of repeat purchases. The social campaigns? Almost zero direct impact on revenue. Lucas: That's the classic correlation versus causation trap. The social posts correlated with sales because they ran during peak shopping seasons, not because they caused anything. The emails were the actual workhorse, but they looked weak because the metrics were vanity metrics. Luna: Right. And the CMO almost made a six-figure mistake. So what is causal AI, really? I mean, beyond the buzz. Lucas: At its core, it's a set of statistical techniques — instrumental variables, difference-in-differences, synthetic controls — that let you actually estimate cause and effect from observational data. The 'AI' part is the automation layer that makes it feasible to run these models continuously, not just once a year. Luna: So it's not magic, it's econometrics on steroids. Lucas: Exactly. The key insight is the counterfactual: what would have happened if you hadn't run that campaign? Traditional attribution just takes credit for everything that happened after the ad. Causal methods build a model of what would have happened anyway — the baseline — and then measure the lift above that. Luna: And that lift is the true incrementality. The hard part is building that counterfactual, right? You can't just rewind time. Lucas: Right. That's why you need experiments. Randomized holdouts are the gold standard. But the trick is that casual AI can get you most of the way there with observational data, if you're careful about confounders. Luna: Like seasonality, or the fact that you're targeting your most loyal customers first. Lucas: Exactly. If your email list is all your best customers, of course they're going to buy more. That doesn't mean the email caused it. You need to compare them to a control group that didn't get the email, but is otherwise similar. Luna: So this is where the data science team becomes essential. You can't just run this in a spreadsheet. Lucas: No, and that's a real barrier for a lot of marketing orgs. But the companies that are doing it well — I'm thinking of brands like Stitch Fix, or the big consumer packaged goods players — they've embedded data scientists right into the marketing team. They're not in some far-off analytics center of excellence. Luna: That makes sense. You need someone who understands both the causal methods and the business context. Otherwise you get a model that's technically perfect but answers the wrong question. Lucas: And there are tools emerging now that make this more accessible. I've seen platforms that let you run a casual analysis in a few clicks, with proper Bayesian priors and all that. But you still need to know what you're looking for. Luna: So what's the first step for a CMO who wants to try this? I mean, most of them don't have a PhD in econometrics. Lucas: Start small. Pick one campaign type — maybe your loyalty emails or a specific paid social channel — and set up a simple holdout test. Just split your audience randomly, exclude ten percent from the campaign, and track the difference in conversion rate over a few weeks. That alone will tell you more than a year of attribution reports. Luna: And that's incremental, right? You don't need to overhaul your whole measurement stack overnight. Lucas: Exactly. The key is to build a culture of experimentation. Once you see the truth in one channel, you start asking the same question everywhere. Luna: And that's when you start finding those hidden growth levers. The thing with the athletic brand — they ended up shifting like twenty percent of their social budget into email and saw a measurable lift in repeat purchase rate within two quarters. Lucas: Which is exactly the kind of result that gets a CMO a seat at the strategy table. When you can show that you've found a high-return, low-cost channel that everyone else is ignoring, that's not just marketing — that's business strategy. Luna: Yeah. And I think there's a broader implication here. We're moving away from this era of 'more data is always better' to 'better questions are what matter.' Causal AI forces you to be explicit about your assumptions, and that's a discipline that pays off everywhere. Lucas: Totally. And if this conversation has sparked something you could actually use in your own planning, I'd love to keep bringing you more of these practical breakdowns. If you want to help keep the show ad-free, you can support us at buy me a coffee dot com slash fexingo. No pressure, but it genuinely helps keep these episodes coming. Luna: Yeah, every little bit goes toward research and guest time. And we really appreciate it. Lucas: So, back to the practical side. I mentioned holdout tests, but let's talk about something a little more advanced: synthetic controls. This is a technique that's been around in academic economics for a while, but it's now filtering into marketing. Luna: Okay, I know the term, but refresh my memory on how it works. Lucas: So imagine you're launching a new product line, and you want to know what it does to overall brand sales. You can't just compare regions with and without the launch, because the regions are different. Synthetic control builds a 'fake' region by weighting a combination of other regions that, together, closely match the pre-launch behavior of your launch region. Then you compare the actual sales in the launch region against that synthetic baseline. Luna: Ah, so it's like a virtual control group. That's clever. Lucas: Right. And it's particularly useful for national campaigns where you can't do a clean geographic split. You can build a synthetic control from a mix of other markets or even from time periods if you have enough data. Luna: So what's the catch? Why isn't every CMO doing this? Lucas: It requires a lot of data and a fair bit of statistical expertise. And the results can be sensitive to how you choose your donor pool — the regions you're weighting. If your model picks a bad donor, you get garbage. But the tools are getting better at flagging that. Luna: So it's a skill you need to build in-house, or hire for. Lucas: Yeah. And I think the CMOs who are really ahead on this are treating it as a core competency, not a one-off experiment. They're building teams that can run these analyses continuously and translate the results into budget decisions. Luna: That's the long game. But for someone just starting, the holdout test is the way in. Lucas: Absolutely. And I'd add: start with a channel that has a clear, short-term conversion goal — like a direct response email or a paid search campaign. Those are easier to measure cleanly than, say, a brand awareness push on TV. Luna: Right, because with TV you'd need a much longer time horizon and a lot more data to see any causal impact. Lucas: Exactly. The principle is the same, but the signal to noise ratio is different. So start with the low-hanging fruit, prove the method works, then scale up. Luna: So, in the end, it's less about the algorithm and more about the discipline of asking, 'What would have happened anyway?' Lucas: That's the heart of it. And the CMOs who embrace that question are going to be the ones who survive the next downturn, because they'll know exactly where every dollar works hardest. Luna: And that's a competitive advantage that no one can copy overnight. Lucas: Right. So if you're a marketing leader, maybe it's time to start that conversation with your data science team. The tools are there, the talent is out there, and the payoff is real. I'd love to hear how it goes.