When most people think about marketing, they think about creativity — great copy, beautiful design, the right message at the right moment. That’s still important. But increasingly, the highest-leverage marketing decisions are engineering decisions.
I’ve spent the last few years building marketing systems for Meituan’s ticket vertical. Here’s how I think about it.
The fundamental problem
Traditional marketing asks: what message should we send?
Algorithm-driven marketing asks: given a budget, a user population, and a business objective, what is the optimal allocation of resources across users, channels, and time?
That’s a fundamentally different question. It’s an optimization problem. And optimization problems have structure you can exploit.
The three layers
I think about marketing systems in three layers:
Layer 1: Targeting — Who should receive this intervention?
Not all users respond equally to marketing. Some users would have converted anyway (you’re wasting money). Some users will never convert no matter what you do (you’re also wasting money). The users you want are the ones who wouldn’t convert without the intervention but would convert with it.
This is the core insight of uplift modeling. You’re not trying to predict who will convert — you’re trying to predict who will convert because of your intervention. That’s a causal question, not a predictive one.
Layer 2: Offer design — What should we offer?
Given that we’ve identified a user worth targeting, what’s the right offer? Too small and it doesn’t move behavior. Too large and you’re leaving money on the table. The optimal offer is the minimum effective dose.
This is where coupon design, discount depth, and offer mechanics come in. Different user segments have different price elasticities. A power user might need a 10% discount; a lapsed user might need 30%.
Layer 3: Budget allocation — How do we distribute resources across the population?
Given a fixed budget, how do we allocate it to maximize total incremental value? This is a constrained optimization problem. The solution involves Lagrangian duality, marginal ROI curves, and real-time bidding.
The causal inference problem
The hardest part of all of this is causality. You can observe who received a coupon and who converted. But you can’t observe the counterfactual — what would have happened if they hadn’t received the coupon?
This is why A/B testing is so important. But A/B testing has limits: it’s slow, it’s expensive, and it can’t tell you why something worked.
More sophisticated approaches use causal forests, double machine learning, and instrumental variables to estimate treatment effects from observational data. These methods let you learn faster and generalize better.
What I’ve learned
Data quality beats model sophistication. A simple model with clean data outperforms a complex model with noisy data. Every time.
The feedback loop is the product. The algorithm is only as good as the data it learns from. Designing the feedback loop — what signals to collect, how to attribute outcomes, how to handle selection bias — is more important than the model architecture.
Business constraints are features, not bugs. Real marketing systems have constraints: budget limits, fairness requirements, brand guidelines. These constraints often encode important business knowledge. Don’t fight them — incorporate them into the optimization.
Measure incrementality, not conversion. The metric that matters is incremental GMV — the revenue you generated because of the marketing, not the revenue that happened to occur after it. This is harder to measure but much more honest.
Marketing is becoming an engineering discipline. The PMs who understand this will build the most effective systems.