Definition of the semantic layer
Between the tables of a data warehouse and the dashboard that displays a figure, a calculation rule applies: which rows to count, over what period, with what exclusions. The semantic layer is the place where this rule is written once, under a business name, for all analysis tools. It holds the shared definitions of metrics, also called measures, and of dimensions, the axes along which a metric breaks down: period, product, region.
The notion is one of common usage: tool documentation describes it, each for its own product. The dbt documentation (docs.getdbt.com, page updated on 29 September 2026, consulted on 12 November 2026) presents its Semantic Layer as the means of defining metrics on top of existing models, outside the reporting layer; the Cube documentation (docs.cube.dev, consulted the same day), as a layer that centralises metric definitions, joins and access rules upstream of each tool.
Content of the layer and possible locations
From one tool to another, under varying names, it brings together:
- metrics, with their aggregation and filters;
- dimensions, including time;
- the relationships between tables, which avoid rewriting joins in each report;
- access rules, common to all tools;
- interfaces (SQL, API) through which a tool requests a metric by a dimension.
The layer can live in the reporting tool: LookML is, according to the Looker documentation, the language of its semantic data models, and Power BI calls the report-ready source a “semantic model”. It can also form an independent component, such as the dbt Semantic Layer or Cube, placed in a modern data stack between transformation and reporting.
A report author chooses a metric there instead of writing a query, which supports self-service BI. An assistant based on a language model can do the same, a use that the dbt and Cube documentation contrast with querying raw tables directly.
Three counts of active subscribers: an example
Take a press group that tracks its digital subscribers. For marketing, an “active subscriber” is any account whose subscription is running. Finance excludes trial offers. The newsroom only counts accounts that logged in during the month. Three dashboards display three totals.
Once the disagreement is settled, the semantic layer receives a metric “paying active subscribers”, with its rule, and two neighbouring metrics under other names. The executive dashboard, the management control spreadsheet and the assistant queried in natural language receive the same total as soon as they request this metric: none of them recalculates the rule.
Semantic layer, data model, data catalogue, OLAP cube: not to be confused
The data model describes tables, their columns and their keys. The semantic layer builds on it and adds named calculations.
The data catalogue inventories datasets and their metadata, without calculating anything. A request sent to the semantic layer returns a result.
The OLAP cube is, according to the Analysis Services documentation, a multidimensional structure made of dimensions and measures, in which each cell contains an aggregated value; Apache Kylin also presents itself as an OLAP engine. LookML and the dbt Semantic Layer, for their part, translate the request into an SQL query executed by the warehouse. The boundary remains porous: Power BI semantic models rely, according to their documentation, on Analysis Services modelling technologies.
The limits: arbitration, wrong data, bypassing
It arbitrates no definition: without a decision from a business owner, it records two competing metrics, and this arbitration falls under data governance. It does not correct wrong data at the source, and a tool that reads the tables directly bypasses it.
Where should this layer live, which model should it rely on, who validates a metric? These three decisions are covered in our page on business intelligence.