What building ML models taught me about ad targeting

Before I ran my first Meta Ads campaign, I trained machine learning models — the kind you write papers about. An ACM paper on heart-attack prediction, an IEEE paper on breast-cancer prediction. I thought I was leaving that world behind when I moved into marketing.

I wasn't. Ad platforms are ML systems, and treating them like one is the single biggest edge I have.

Your account structure is your training data

A model is only as good as the data you feed it. When you cram cold prospecting and warm retargeting into one ad set, you're handing the algorithm mislabeled data — it can't learn who your customer is because you've mixed two different questions into one dataset.

Optimization events are loss functions

In ML, you choose what the model minimizes. Optimize a campaign for clicks and you'll get clicks — from people who click everything. Optimize for qualified leads and the system hunts for those instead. Most "Meta Ads doesn't work" stories are loss-function stories.

Small budgets overfit

A model trained on 50 examples memorizes instead of learning. An ad set that gets 5 conversions a week never leaves the learning phase — its decisions are noise dressed up as strategy. Consolidate until each ad set gets enough signal, or you're paying for randomness.

The machines running your campaigns don't care about creative awards. Feed them clean data, point them at the right objective, give them enough signal — the same three rules that make a research model publishable.