AI Win-Rate: A Board Metric for AI-Mediated Buying (Methodology v2.0)

Abstract

In sales, win rate is opportunities won divided by opportunities available. It is one of the few commercial measures that survives contact with a board. In AI-mediated buying it has not been measurable, because the pipeline that produces it does not exist: nothing is logged, losses are invisible, and the outcomes have no equivalent in a sales funnel. This document defines AI Win-Rate℠ — the proportion of decision conversations in which an AI assistant recommends a brand, product or service outright. Put plainly: who wins the final recommendation in a multi-turn conversation. New in version 2.0. Version 1.0 recorded that the width at which a figure ceases to be a rate and becomes a range was a convention requiring a stated basis. This version supplies that basis as a procedure any instrument applies to itself: the three steps by which an instrument establishes its own precision floor, and a reporting threshold — published as a rate where the 95% interval is no wider than ±10 points, as a range between ±10 and ±20, and withheld beyond ±20 — expressed in decision conversations rather than prompts, queries or API calls. Also new: a paired requirement for before-and-after claims, which detects a genuine ten-point improvement four times in five where an unpaired comparison misses it in ninety-three cases out of a hundred; a requirement that a scenario set declare the basis on which it was constructed, since fixing a set in advance does not on its own establish that it represents anything; the separation of measurement from commercial modelling; how figures from different assistants combine; and the statement that replication is not repetition over time, longitudinal smoothing being unable to distinguish model drift from inference noise. The win rate is terminal-only. Earlier-turn leadership and displacement travel beside it as declared companions, never as components of it. A first answer characteristically does not resolve to a single entity and therefore contains no win to count; any figure derived from position in an opening list is a prediction of a win rate calibrated against one, and is not a win rate. Supersedes version 1.0 (DOI 10.5281/zenodo.22280231).