Many AI visibility tools publish per-dimension confidence scores alongside the main 0–100 scores. The confidence number typically indicates how consistent or certain the model was when generating the answer. Used correctly, it is a genuinely useful signal — it helps separate stable findings from noisy ones. Used incorrectly, it is worse than useless. It can lead a team to trust a high-confidence-but-wrong answer and dismiss a low-confidence-but-correct one. This post unpacks what the confidence score actually measures, how to read it alongside the main score, and — importantly — when to ignore it.
In a ranked list, the unit of success is position. You are first, or third, or eleventh. In an AI answer, there is no list. There is a paragraph. Your brand either appears inside the paragraph — cited, named, described — or it does not. Citation has quietly replaced ranking as the metric that matters, and the replacement changes how you work. Link-building was a decades-long craft built around one unit. Citation-building is a parallel craft built around a different one, and the distinction matters.
Share of Voice has been a marketing fixture for thirty years. It measured your brand's share of media mentions, press coverage, or paid impressions against competitors. It was crude, it was useful, and it gave boards a number to argue over. The underlying channel has shifted — media coverage and paid impressions are no longer where most buyers first hear your brand named. The channel that matters most today is the composed answer of a language model, and the right analog for SOV in that channel has a different name: Share of Model.
Every AI brand visibility audit ever run, regardless of vendor or methodology, is trying to answer some combination of three questions. Do the models know you exist? Do they surface you when it matters? Do they describe you accurately? Each question has a different remedy, and tools that collapse them into a single score make it impossible to tell which problem you actually have. This post is a practical frame for interpreting audit results — Recognition, Recall, Reality — and what separates a useful report from a decorative one.
A subtle distinction shapes almost every practical decision in AI brand visibility. There is the brand as the model has learned it — baked into its parameters from training data. And there is the brand as the model describes it in a specific answer, shaped by retrieval, the user's question, the conversation history, and post-processing. The first is memory. The second is context. Conflating the two is how teams end up fixing the wrong thing. The distinction is simple once you name it, and useful once you use it.
