How the number is made, and what it does not prove

Answer engines are probabilistic: phrasing shifts the answer, history shifts the answer, a model update shifts the whole league overnight. One query run is an anecdote. So we sample.

• Each race generates 8 to 15 phrasing variants of a real buyer question, approved before they go live.
• Paid census: 25 runs per race per engine per month, distributed across days and times, across all five engines (125 runs per race). The free scan runs 5 per engine, labeled a low-confidence baseline.
• Runs execute from clean sessions with no account history by default. Personalization is an acknowledged, unmeasurable residual.

Every proportion we show carries its sample size and a 95% confidence interval (Wilson, for small samples). Below the significance floor we show "census in progress," never a number dressed up as certainty. A 12% share of voice with no denominator is marketing, not measurement.

The Authority Graph Score compresses recommendation readiness into one number from 0 to 100. It is a proprietary readiness score, not a number any AI system computes. Today the score itself is driven by one signal we can measure cleanly across repeated live races: your recommended-or-considered share across your races. Alongside it we report the corroboration (citation and trust) behind those answers as a measured driver, shown next to the number rather than folded into it. The race shares carry their run count and a 95% confidence interval; the corroboration driver is a measured rate shown over its run count. We do not score what we cannot yet instrument, and no number is shown without the denominator it was measured over.

The full model has eight drivers. They are the framework the score grows into, not eight separately instrumented inputs today. As each driver's signal becomes measurable, it joins the score, and the decomposition below arrives in a later phase. We name all eight here so you can see where the score is headed:

• It is not a number any AI system computes. It is our model of the evidence the engines weigh, sampled across repeated live races.
• It does not claim a causal link to revenue. No tool in this category has shown a consistent mentions-to-traffic correlation, and we will not pretend otherwise. We sell the evidence record as an asset, not a traffic promise.
• The goal is not 100, or even "high." It is to outscore the names winning the races that actually pay you.

Here is the why behind the score. The evidence record is the diligence-verifiable asset: years-deep, machine-readable third-party corroboration that sits in public where a buyer can check it. Recommendation equity is the reason that record matters, since acquirers underwrite durable demand that does not depend on the founder, and durable demand tends to show up in the multiple a buyer pays. That is the thesis, not a promise the score makes. AEO Radar measures recommendation readiness and surfaces the evidence record. It does not guarantee a sale, a valuation, or a buyer, and any multiple language is illustrative.