Educational guide
1. Confidence score is not one universal metric
The label can refer to a classifier probability, a retrieval score, a token-level probability, an evaluator result or a product-specific composite. Before interpreting a number, find its documented definition, scale, unit of analysis and calculation version. If those are missing, call it an undocumented score rather than model confidence.2. A score is not ground truth
A value such as 0.8 is a model or product output under defined conditions, not proof that a statement is 80% true. Google's classification guidance explicitly distinguishes a predicted probability from reality and shows how thresholds create false positives and false negatives: Thresholds and the confusion matrix.3. Calibration determines what probability-like scores mean
A probability-like score is useful only if its relationship to observed outcomes has been evaluated on representative data. A calibrated 0.8 bucket should correspond to the defined outcome at roughly that rate over many relevant examples. Calibration can drift when data, models, prompts or product logic change.4. Use the SCALE interpretation checklist
- S — Source: Which model, product or evaluator produced the score?
- C — Calculation: Is it a probability, similarity, rubric result or weighted composite?
- A — Assessment: On what dataset and ground truth was it tested or calibrated?
- L — Level: Does it describe a token, claim, answer, source, prompt set or account?
- E — Effective date: Which version and collection period does it represent?
5. Worked comparison
System A reports 0.80 as a calibrated probability that a document belongs to a category. System B reports 80 as a weighted content-review score. The numbers look similar but cannot be compared: they use different calculations, units and validation. First interpret each against its own documentation and observed outcomes.6. What this means for AI visibility
Do not assume that answer systems expose a stable confidence score that websites can boost with schema, wording or backlinks. Measure observable outcomes instead: whether a brand was mentioned, how it was framed, which sources were exposed, whether facts were correct and how those observations changed across a fixed prompt set. The KNWN Visibility measurement guide documents that evidence-first approach.Interpretation note
AI and search behavior changes over time. Guidance does not promise discovery, ranking, citations, traffic or commercial outcomes; confirm platform-specific details against primary documentation.