Can AI Fix Advertising’s Measurement Problem?
Advertising’s measurement problem has never been a shortage of models. It is that the effect worth measuring happens inside people and cannot be observed directly, so every method is an inference dressed as a number.
Attribution, mix modelling, incrementality testing and brand tracking are all attempts at the same impossible reading. Each is defensible, none is complete, and the industry has spent thirty years choosing between their different failure modes.
What AI Genuinely Improves
Real gains exist and are worth having. Models can handle far more variables, cope with sparser data, and run continuously rather than quarterly. Mix modelling that once took a specialist team a month can now run often enough to inform decisions while they are still open.
That is a meaningful change in tempo. It is not a change in kind.
Why The Measurement Problem Survives
A more sophisticated model still cannot see the thing it is estimating. It infers persuasion from behaviour, and behaviour is shaped by price, distribution, weather, competitors and a hundred things nobody logged.
A model that produces a confident number from data that never contained the answer has not solved the problem. It has hidden it better.
The Risk That Comes With The Improvement
The danger is not inaccuracy, it is unfalsifiability. An older model was crude and its assumptions were legible, so a marketer could argue with them. A model with a thousand inputs produces an output nobody in the room can interrogate, and it tends to win arguments on presentation rather than merit.
The useful discipline is unchanged: ask what would have to be true for the number to be wrong, and check whether anyone in the room can answer. A model nobody can argue with is not a better model, it is an unaccountable one, and marketing has bought several of those before. See also how media plans are changing, where the same inference is already being used to buy.