--- language: en title: "Factual accuracy score" description: "A factual accuracy score compares verifiable claims against defined references. The report should show errors and inconclusive cases as well as the percentage." canonical: "https://tropk.ai/en/resources/glossary/factual-accuracy-score" last_modified: "2026-10-12T16:12:07.967Z" --- # Factual accuracy score Reviewed on 2026-10-06. ## What it means A factual accuracy score summarizes how many verifiable claims in [AI responses](/en/resources/glossary/ai-response) match a defined reference. The calculation depends on how claims are selected, the evidence used, and the evaluation rules. In Profound’s help documentation consulted in October 2026, Accuracy Score divides correct claims by the total of correct and incorrect claims, excluding unverifiable ones. ## Why it matters for marketing For marketing, this metric helps distinguish incorrect factual descriptions from unfavorable opinions. Stating that geographic coverage exists where it does not requires a different review from calling a service expensive. The team can prioritize errors about service terms and check the corresponding public documentation. A high percentage does not eliminate the risk of an important error hidden among many simple claims. ## Example in practice Imagine a car rental company whose hypothetical audit extracts 25 claims about vehicle pickup. The current policy lets the team verify 20: 18 are correct, and two claim that pickup is possible without ID. Under the stated protocol, the score is 90%. The five inconclusive claims, 20% of the extracted set, appear separately in the report. The team prioritizes the information about documentation, even with an apparently favorable overall score. ## How to apply or verify Save the response, the isolated claim, and the reference, including its date, version, and applicable conditions. Classify each case as correct, incorrect, or inconclusive, with a rationale. Check whether the reference covers the same region and service type. If you exclude inconclusive cases from the denominator, report their number and proportion of the extracted set; lack of evidence should not be treated as an error. ## Important distinctions Factual accuracy here means whether claims are correct, not a classifier’s statistical precision or the confidence stated by the model. The reference may also be outdated or incomplete. Compare scores only when the rules and data sources are compatible. Responses with no evaluable claims are not automatically assigned a score of zero or one hundred: they are left unscored, with the reason recorded. ## Sources and references - [The AI Search Glossary: A definition of 300+ terms]() — Profound - [Profound glossary]() — Profound ## Continue a leitura - [AI hallucination]() - [Profound’s FactCheck]() - [Factual accuracy prompt]() ## About the author [Luiz Gardelin]() Luiz Gardelin is the founder of Tropk.ai and works in GEO/AEO, product vision, strategy, and growth, helping brands understand and improve their presence in AI responses. --- [All pages in Markdown]()