Latest / The CMO Podcast with Fexingo: Marketing Leadership, Budgets, and Executive Strategy / How CMOs Are Using AI for Predictive Customer Lifetime Value
Transcript
- Lucas: Luna, I want to talk about something that's quietly reshaping how CMOs decide where to spend their next dollar — predictive customer lifetime value. Not the old RFM spreadsheet approach, but actual machine-learning models that estimate what a person will be worth before they've even made their second purchase. Luna: I've seen some brands claim they're doing this, but I suspect the reality is messier. How many are actually pulling it off? Lucas: Less than twenty percent, according to a study I read from the Marketing Science Institute earlier this year. But the ones who do it well are seeing real shifts. Take a DTC brand in the home-goods space — call them a mid-market player with maybe fifty million in annual revenue. They started feeding their customer database into a neural network that looked at three hundred behavioral signals per visitor. Luna: Three hundred — that's a lot. What kind of signals are we talking about? Lucas: Time on site, obviously. But also scroll depth on product pages, how many times they visited before purchasing, whether they opened the welcome email within an hour, what device they used, even the hour of day they first arrived. The model learned which combinations predicted a customer who'd still be buying twelve months later versus someone who'd churn after one order. Luna: And then they used that to decide which prospects to bid on in acquisition channels? Lucas: Exactly. They scored every new email signup against the predicted LTV model and then allocated ad spend accordingly. High-scoring prospects got retargeted with higher frequency and higher-value creative. Low-scoring prospects got a single touch and then dropped. Within six months, their blended customer acquisition cost dropped twenty-two percent. Luna: Twenty-two percent — that's meaningful. But I wonder about the data prerequisite. Most brands don't have a clean, long enough history to train a model like that. Lucas: No, you're right. The general rule of thumb I've heard from data scientists is you need at least five thousand completed customer journeys — someone who entered, purchased, and either churned or stayed — ideally with twelve months of transaction data. Below that, the model's just overfitting to noise. A lot of younger brands hit a wall at that point. Luna: So it's a scale game. But for brands that have that history, what's the biggest internal friction? Lucas: Finance. CFOs are used to looking at trailing metrics — what already happened. A predicted LTV is a forecast, it's probabilistic, and finance teams hate allocating budget based on a probability distribution. So the CMO has to build a governance framework where the model's predictions are validated quarterly against actual retention curves. Luna: Meaning you let the model run for a quarter, then compare its predictions to what really happened, and adjust the model before the next budget cycle. Lucas: Exactly. And that's actually a cultural win — it makes marketing accountable to a forward-looking metric instead of just 'we spent this much and got this many leads.' Luna: I think the biggest shift here is that it forces CMOs to stop obsessing over CAC in isolation. You can have a low CAC for customers who never buy again and a high CAC for customers who stay for three years — and the high CAC is actually the better investment. Lucas: Right. That's the unit-economics mindset that venture-backed DTC brands have been evangelizing for years, but now it's migrating to more traditional CPG and retail companies. I talked to the head of marketing analytics at a large beverage company — they're piloting a similar model with their loyalty program data. Luna: What signals are they using? Lucas: They're pulling in in-store purchase frequency via loyalty card scans, but also digital engagement like recipe page views and newsletter clicks. The model assigns a predicted LTV score to every loyalty member, and then the marketing team tailors the coupon frequency — heavy discount offers only go to segments predicted to be high ltv, because they know those customers will stick around and the discount pays back. Luna: That's smart. low ltv segments might still get offers, but probably smaller ones or less frequent. Lucas: And that's good for margin. You're not burning promotional dollars on people who were going to leave anyway. One of the CMOs I spoke with said this model effectively changed their promotion strategy from 'one-size-fits-all' to 'graduated precision.' Luna: Which sounds obvious in retrospect, but most companies don't have the data infrastructure to execute it. Lucas: No, and that's the real barrier — it's not the algorithm, it's the data integration. You need to stitch together web analytics, CRM, point of sale, email platforms, maybe even call center logs. That's a six-month project for a typical mid-market brand, and it requires a data engineer who understands marketing data, not just backend systems. Luna: So the CMO has to be the one pushing for that integration. It can't just be an IT project. Lucas: Exactly. The CMO has to own the use case and articulate the ROI of the data unification. Without that, the model never gets built because it's always deprioritized. Luna: I want to talk about the inevitable skepticism — when you present a predicted LTV number to a room of executives, someone always asks 'how do you know it's accurate?' Lucas: And the honest answer is: you don't, not in the moment. But you can build confidence through backtesting. Take your historical data, pretend it's a year ago, run the model on that snapshot, and see how well it predicted the actual LTV of customers who signed up then. If you get a mean absolute error under fifteen percent, you have a defensible model. Luna: Fifteen percent — that's a good benchmark. I've heard that some brands are now feeding predicted LTV directly into their media-bidding algorithms. So instead of a flat CPA target, the bid adjusts per impression based on the predicted LTV of the cookie or ID. Lucas: That's the next frontier. Programmatic platforms are starting to offer custom machine-learning model integrations. You upload your predicted LTV scores, and the platform optimizes bids toward that metric instead of clicks or conversions. It's basically turning your customer data into a direct input for ad auction strategy. Luna: That feels like a major shift. For decades marketing attribution has been about 'what caused the sale.' Now it's 'what kind of customer will this sale produce.' Lucas: That's a really clean way to put it. And it changes the conversation with the CFO from 'we spent X and got Y customers' to 'we spent X and acquired a cohort with a predicted Z lifetime value, which implies a return on investment of W percent.' That's a much stronger narrative for budget protection. Luna: Speaking of the CFO — I know this is a bit of a pivot, but it ties back to how valuable these kinds of data-driven conversations are. If today's episode gave you something you can actually use — a framework, a benchmark, a way to talk to your finance team about LTV — then it might be worth the price of a coffee to you. Lucas: Yeah, look, we keep this show ad-free by design, and listener support is what makes that possible. If you found value in the specifics we covered today, that's the link — buy me a coffee dot com slash fexingo. No pressure, just a way to keep the conversations going. Luna: And every bit of that support genuinely helps us bring on more practitioners and dig into the details that actually move the needle. So thank you for considering it. Lucas: Alright, back to the playbook. One nuance I want to hit — the model doesn't just predict total LTV. It can predict the shape of the revenue curve. Some customers have a slow ramp — they buy small, then gradually increase. Others have a spike at month three and then decline. Knowing the shape changes how you schedule your retention spend. Luna: So if you know a segment tends to spike at month three, you'd put a retention campaign at month two and a half, not at month one. Lucas: Precisely. And that's where the model moves from being a budgeting tool to an orchestration tool. You're not just deciding how much to spend — you're deciding when to spend it, on what channel, with what message. That's the difference between a tactical CLV score and a strategic one. Luna: I think that's the takeaway a lot of our listeners need to hear. It's not about buying the fanciest AI platform. It's about getting your data clean, setting up the feedback loop, and then using the output to make one better decision per quarter. Lucas: Exactly. Start with a question — not with technology. The question is: if I could predict which new customers will be my best customers in eighteen months, how would I change my acquisition strategy today? Answer that, and the model becomes a means, not an end. Luna: And maybe the best place to start is with a simple regression on your existing data before you even talk about neural networks. The incremental lift from moving from a spreadsheet to a basic model is often bigger than the lift from moving from a basic model to a deep learning one. Lucas: That's a really important point. Don't let perfect be the enemy of good. A logistic regression that uses just purchase frequency and average order value will already outperform RFM for most brands. You can layer in more complexity as you go. Luna: I like that — predictive LTV as an iterative practice, not a one-time software implementation. Lucas: That's the right way to think about it. And if you start today with a simple model, by this time next year you'll have the data to build the better one.