How it works
Two numbers decide almost everything.
What a customer is worth to you, and what a customer costs to acquire. Everything Jules does is arranged around comparing those two honestly — and refusing when the comparison does not work.
The ceiling: what you can afford
A customer’s lifetime value multiplied by your contribution margin — the profit actually available to spend on winning them. If you do not know your lifetime value, Jules asks for a maximum you are willing to pay instead. It never guesses, because an invented number here quietly invalidates every figure downstream.
The floor: what a customer costs
Cost per click divided by conversion rate, for your market, your industry and the specific kind of conversion you are selling — a booking is not a download and a B2B licence is not an impulse purchase. Benchmarks are held at the finest detail available and widened when the data is thin, and the answer says how far it had to widen.
There is a second step most tools skip. What an ad platform can measure is a form filled in or a button clicked; what you actually care about is a customer. The rate between those two is usually offline and almost always ignored, which is how a plan can look comfortable by a factor of ten and be wrong. Jules models it explicitly, and if you do not know it, says so rather than assuming it away.
The verdict, in bands
Comparing a single pessimistic ceiling against a single pessimistic floor would refuse almost everybody, so the comparison is banded. Comfortably above the range is viable; inside it is marginal and worth a test before commitment; below it is unviable, and no amount of clever optimisation closes that gap.
An unviable verdict produced by arithmetic is final. Nothing further in the system is permitted to move a channel off it — not the strategist, not the language model, not a persuasive prompt.
Then it designs the tests
Each plan carries numbered assumptions, each with what happens if it turns out to be false and a test that would find out. Tests are ordered by what they teach divided by what they cost, so the free desk research runs before the expensive experiment. Each one is registered with its success threshold before it starts, so the bar cannot move afterwards.
If your budget cannot buy enough data to detect the effect you are looking for, the test is labelled a directional probe everywhere it appears, rather than being quietly reported as a result.
Where the model is allowed to speak
After all of it. The language model writes the explanation around numbers it has already been handed, and cannot alter one. It runs on open weights pinned to an exact revision, so a strategy produced in March can be reproduced in September rather than merely remembered.