A business description of data: what measures mean, which records relate, and how figures are grouped and computed for an answer.
A semantic model turns fields such as amount, credit type, and posting date into agreed concepts such as net revenue for September. It tells the person or tool using the data what a measure includes, how records connect, and what level of detail an answer can support.
The terms overlap in practice. A data model describes structure and relationships. A semantic model adds business meaning and calculation rules. A semantic layer makes shared definitions available to reporting and analytical tools. The important check is whether those definitions are actually applied when an answer is computed.
A semantic model is a set of definitions, rather than one formula. For each measure, agree its inputs, inclusions, exclusions, period, aggregation, and row level. Name the relationships that allow the measure to be grouped by customer, cost center, or product, and identify who approves changes.
Net invoiced revenue = gross invoiced revenue − issued credits, within the same agreed scope and posting period.
That illustrative definition differs from bookings, cash receipts, or a revenue measure with other exclusions. The business owner must choose the basis that fits the report. Writing a familiar label over an amount column does not settle it.
A US finance team agrees to report posted invoices less issued credits, in USD, by customer, for September. Each invoice and credit appears once and maps to one customer.
| Customer | Gross invoiced revenue | Issued credits | Net invoiced revenue |
|---|---|---|---|
| A | $120,000 | $10,000 | $110,000 |
| B | $80,000 | $5,000 | $75,000 |
| Total | $200,000 | $15,000 | $185,000 |
The semantic model establishes why the answer is $185,000: the selected definition includes both customers, uses posting month, and deducts issued credits. Its relationships should preserve the total when the result is grouped by customer. The same question over the same records and definition should produce the same figure.
A label hides a policy choice. One team means invoiced revenue, another means bookings. Agree the measure before comparing the answers.
Relationships duplicate amounts. A customer mapping with multiple matches can repeat the same invoice. Check keys, unmatched records, and the total after grouping.
Definitions change without a version. An old board pack and a new answer can use different rules. Keep the effective date and the version used by each answer; see versioned metric definitions for the full method.
For business teams working from files, Covirage applies versioned definitions through its calculation tools without requiring a warehouse or a data team. The tools compute every figure; the external AI model explains the results. See metrics governance to apply this to a recurring report.