Agree the meaning first
When revenue means three things in three places, a person or a tool picks one at random. Getting stakeholders to sign each definition is a business decision. No build can make it for you.
What the evidence says about definitions
A vendor comparison found questions answered through a semantic layer were right more often than text written straight against tables: 98.2% and 100.0% against 90.0% and 84.1% on modelled data, and 72.7% against 64.5% on the original schema, where it could not answer questions needing too many joins. A separate preprint found a short file of written definitions lifted accuracy by 17 to 23 points. Both are small, specific tests, not a promise about your data.
Control totals are the test
Each metric should reconcile to a figure someone trusts, within a tolerance written down. Questions outside the definitions should be refused, not guessed. Ask for a held-back set of questions you wrote.
- A control total per metric.
- Ten held-back questions with expected values.
- Out-of-scope questions that return an error.
Which job fits
The core-metrics job is the start: three metrics from £495, ten quoted. A data dictionary and a one-report lineage map help where nobody can say what a column means or where a number comes from. A question tool over approved views comes after the definitions, not before.
Sources and limits
- dbt: semantic layer vs text-to-SQL, 2026 Checked 2026-10-12.
- On modelled data text-to-SQL scored 90.0% and 84.1% for two models against 98.2% and 100.0% for the semantic layer, and on the original schema 64.5% against 72.7%; the post says the semantic layer answers only what is modelled and returns an error when it cannot answer.
- arXiv 2604.25149: Semantic Layers for Reliable LLM-Powered Data Analytics Checked 2026-10-12.
- On 100 questions and three frontier models, a short hand-written markdown file of definitions improved accuracy by 17 to 23 percentage points, from 45.5-50.5% to 67.7-68.7%.