Persuasion can be measured. The market already keeps the score.
Advertising is the largest running experiment in human persuasion. Every day, brands put different messages for the same product in front of real people and pay to learn which one works. The answer comes back as behaviour: who stopped, who watched, who paid. Few other fields give persuasion a ground truth like this. It is not a judge’s opinion. It is what people did.
Can a model tell which message people will act on?
Show an expert marketer two ads for the same product and they can often tell, in seconds, which one will sell. They hear when a voice sounds sure and when it sounds like a script. They see when an opening earns the next second and when it asks for it. This instinct takes years to build, and it is rarely written down.
Frontier models can describe an ad in detail: the hook, the offer, the emotion it aims for. They still pick the one that lost. Knowing the parts of a message is not the same as knowing how a person will respond to it.
Can a model write a message people act on?
Telling a good message apart is half the skill. The other half is making one: a script, an image or a line of copy that a stranger who never asked to hear from you stops for, believes, and acts on.
There is no answer key. The same words land in one market and fail in the next. What worked last month is tired this month. A model has to learn what real people want, not what a rubric rewards.
Can a model spend money where it works?
Marketing is not one message. It is a long series of decisions made under uncertainty: which ad gets the budget, which one to stop, when a winner starts to wear out. Each choice costs real money, and its result arrives days later, mixed with noise.
This is the kind of long, feedback-driven work that reinforcement learning is built for, and every decision can be checked against what really happened next.
Can it persuade without deceiving anyone?
A model that persuades well is a model that could manipulate well. So we measure both: whether a model moves people, and whether it does so with true claims, fair offers and no pressure. A benchmark that rewards only the outcome would train the wrong thing.
We grade with real outcomes wherever we have them: what was spent, what sold, how long an ad kept running. Where we do not have them, we use expert marketers, and model judges only last. Each result says which one it comes from.
This is early work. Outcomes are noisy: the same ad can win or lose on its timing, its audience or its budget, and much of that sits outside the message itself. We will publish our methods and our limits with every result, so that others can check the work and build on it.