What is machine learning in advertising?

Rules versus learned patterns

A rules-based system does what it was told: bid more on Tuesdays, exclude this placement, cap at this frequency. Every behaviour is written down by someone.

A machine learning system is given examples and objectives and derives the behaviour itself. Nobody writes down that a particular combination of context and time-of-day predicts conversion; the model finds it in the data.

That is more powerful and considerably less inspectable, which is the trade every applied machine learning decision in advertising comes back to.

Where it does the work

• Bid prediction. Estimating the value of an impression — the highest-volume application in the industry.
• Audience expansion. Finding users resembling converters without a human specifying the resemblance.
• Creative selection. Choosing which variant to serve in a given context.
• [Fraud detection](/resources/glossary/mobile-ad-fraud). Classifying anomalous traffic against an adversary that adapts.
• Intent estimation on conversational surfaces — see intent density.

How it fails in this domain

Three failure modes recur often enough to plan around.

Optimising the proxy. A model maximises the objective it is scored on. If that objective is a biased attribution model, the system will spend confidently into the bias, and faster than a human would. This is the single most expensive machine learning failure in advertising, and it looks like success in reporting.

Feedback loops. A model that allocates spend also generates the data it next learns from. Channels it starves stop producing evidence that they work, which confirms the original decision. Deliberate exploration budget is the standard defence.

Distribution shift. A model trained before a market change keeps applying the old relationships. Privacy changes, new surfaces and seasonal breaks all do this, and the model does not announce it.