preciseHow a number gets onto the page.
Three layers, each reviewable: contribution methods value every attribute, decision models turn values into a recommendation, and the record checks the recommendation against what was measured.
This page describes method, not results. Every figure on this site is a synthetic example unless it says otherwise.
1 · Contribution methods
An impression carries many attributes at once: the creative, the audience segment, the publisher, the supply path, the DSP, the daypart, the frequency. Each can contribute value on its own and change value in the company of the others.
Precise assigns each attribute the share of the outcome it can fairly claim, individually and in combination, against the cost it added. The unit is the same everywhere: outcome value per $1 of media spend and fees, so a publisher and a data layer can be read on one scale.
Where two attributes reached the same people, their contributions are not summed. The overlap is named and, where only platform reports exist, the number is labeled gross.
2 · Decision models
A value is not yet a decision. Each combination can absorb only so much spending before the next dollar returns less; the response curve says how much. Precise reads the curve for the cell that would give and the cell that would receive, and prices the move: how much, and what is expected at the horizon.
Every expected figure carries a range. The range comes from scored history: moves of the same kind whose outcomes have already been measured. Fewer scored moves, wider range. Where there is not enough evidence, that is the answer, said in words.
3 · The record
A recommendation is written down with what was known, what was chosen, what was expected, and, later, what happened and what it taught. Sent, reported and observed are three separate facts. Scoring happens at the agreed horizon against a holdout that kept the old setting, so the next range stands on more evidence than the last.
Nothing on the page is typed by a model. Words come from the assistant; every figure comes from these methods with its basis attached.
Scored history for one kind of move: what a predicted line stands on.
Each point is one scored move: the gain in return per $1 at the 7-day horizon against its holdout. Synthetic fixture, n = 14.