A head of analytics at a retailer ships a metrics layer in a quarter. Churn rate is defined once, every dashboard agrees, and the Monday meeting gets shorter. Then Finance asks why churn excludes trial accounts, and the retention team asks why it includes them. The file has a formula. It has no answer. The metrics layer vs semantic layer question lives in that gap, and the difference between the two is a noun.
A metrics layer is built around a measure: revenue, margin, churn. It says how a number is computed and which fields it can be grouped by. A semantic layer, as it should be, is built around a concept: a customer, an order, a policy. It says what the thing is, who agreed that, and which records qualify, and then it holds the calculation over it. Both are governed records. They are about different nouns.
A metrics layer is organized around a measure and says how to compute it. A semantic layer is organized around a concept and says what it is, with a named owner, before any calculation runs. The first is what the market ships. The second is what an AI agent needs.
In our post What Is a Semantic Layer? we set out the four things a layer holds: concepts, metrics, mappings, and entitlements. Most products hold the second and third. The difference between the two layers is which noun sits at the centre of the record.
| Question | Metrics layer | Semantic layer |
|---|---|---|
| The noun at the centre | A measure: revenue, churn, margin | A concept: customer, order, policy |
| What it does with that noun | Computes it once, the same way everywhere | Defines it, names an owner, then attaches the calculations over it |
| What "customer" is in it | A join key with an optional description | A class with a definition, a rule, a scope, and a named owner |
| What a change looks like | A new formula in a file, reviewed in a pull request | A new version of the definition, approved by its owner, the old one still queryable |
Take the file apart. dbt's documentation for metrics (dbt v1.12 and later, read 2026-11-22) lists the parameters of a metric as name, description, type, label, config, and filter. Owner appears in that page's examples as a config key, meta.owner, set to a team handle (dbt documentation). The same vendor's overview of MetricFlow says "Dimensions add context to metrics and without them, a metric is simply a number for all time" (dbt documentation).
| Field | What it holds | What kind of word it is |
|---|---|---|
| type | Simple, ratio, cumulative, or derived, plus an aggregation | The verb: what is done to the column |
| expression | The column, or SQL over it | What the verb is done to, inherited from the table |
| dimensions | The fields it can be grouped by | Adjectives: which slice |
| description | Free text | A sentence nothing tests |
| config.meta.owner | A team handle | A label on the file |
Every row is a verb, an adjective, or a label. The nouns in the formula, the orders summed and the customers counted, arrive with the table name. The product defines the calculation with precision and leaves the noun to whoever built the table.
A metrics layer can define a calculation. It cannot define a concept. It needs defending, because dbt's documentation on entities (last updated 2026-07-20) opens with "Entities are real-world concepts in a business, such as customers, transactions, and ad campaigns". The next sentence says what the product does with them: "In our Semantic Layer models, these entities serve as a join key across semantic models." Name and type are required; the description is optional (dbt documentation). Our post Semantic Layer vs Ontology: Which One Says What a Customer Is shows what the key leaves out.
| A concept needs | In a metrics file | On a governed record |
|---|---|---|
| A definition in words | A description string on the entity | The definition, versioned, with the date it took effect |
| A rule that tests a record | A data test on a column, if someone writes one | A SHACL shape (W3C Shapes Constraint Language), run on every load |
| A scope | One definition per project | Retail lines only; the commercial division's customer is a second record |
| An owner whose approval counts | meta.owner, a key-value pair | A named person, an approval date, and a review date |
The owner row is where dbt's own documentation makes the point. The meta config "accepts any key-value pairs", and "Changes to meta, including at the column level, don't trigger state:modified. dbt treats meta (and tags) as metadata only" (dbt documentation). So the owner of a definition is a label, and changing it is, in the product's own terms, metadata. A metric definition nobody approved is inferred metadata with a formula attached, and our post Ground Truth: SME-Approved vs Inferred Metadata covers what the approval adds.
Yes. We read five competing pages on this query, and every one lands there. Kyvos (undated) says "one component of a broader semantic layer"; Querio (2025-12-31) says "a specific part of the larger semantic framework".
| Page | Its sentence |
|---|---|
| Cube, updated 2026-08-28 | "A metrics layer governs measures. A semantic layer governs measures plus entities, dimensions, join paths, and access rules." |
| A catalog vendor's page on this query, 2026-09-10, updated 2026-09-18 | "Every metrics layer does semantic-layer work at a smaller scale; not every semantic layer is a metrics layer, since most also model things a KPI formula never touches." |
| datus, 2026-06-08, updated 2026-06-10 | "It is the metrics-focused subset of a broader semantic layer." |
True, and it leaves out which component comes first. What the pages say the semantic layer adds, entities, dimensions, join paths, and access rules, is the grammar of the warehouse: how tables connect and who may read them. None of it says what a customer is. Cube's page says a metrics layer "gives the agent certified measures; it doesn't necessarily give it the entity graph needed to slice those measures in a way nobody pre-registered". The catalog vendor's page says a metrics layer cannot guarantee that "customer" means the same thing in billing, CRM, and support "because entity resolution sits outside a metrics layer's stated scope". Each then places its own product where the gap is. The missing component is the concept, and it comes first, because every metric is a calculation over concepts someone defined first.
Here is the retailer's churn rate with its nouns pulled out: customers lost in the period, divided by customers at the start of it. The metrics layer holds the division. Every noun is a decision somebody has to have made.
| Noun in the formula | What the metrics file holds | What has to be decided first | Who decides |
|---|---|---|---|
| Customer | A join key to dim_customer | Does a trial account count? Does a lapsed one? | Finance, with a named owner |
| Lost | A filter on a status code | Cancelled, or ninety days inactive, or both? | The retention lead, who disagrees with Finance |
| The period | A time dimension at month grain | Calendar month or billing cycle, and in which time zone? | Finance, and it changes at year end |
Three nouns, three decisions, two owners who disagree. The metrics layer computes the same disputed number everywhere, which is what the dashboard problem needed. The concept-first order puts each decision on a record with a name on it and attaches the formula to that record. Then the definition and the calculation cannot drift apart. Our post How to Build a Semantic Layer: Platform-Neutral Steps and What Each Native Layer Leaves Out gives that order step by step.
Often. The dashboard problem is real, and a metrics layer is the right tool. Three things change the answer.
Our post Semantic Layer Tools: Metrics Layers, Native Views and Meaning Layers Compared places each product in its category.
Change the definition. Decide that trial accounts no longer count, and watch what moves.
A layer that passes the second test has the concept as its unit of record. A layer that passes only the first is a metrics layer, whatever the box says.
TopQuadrant does not sell the dashboard, the warehouse, or the metrics file. TQ Data Foundation is the governed semantic layer those things read from, and the concept is its unit of record. Every metric carries its business definition, its owner, its calculation, and its mapping to physical columns. The plain-language meaning and the logic live on the same governed asset, so disagreements resolve to a person. Version history shows what the metric meant last quarter and what changed since. When the definition changes, it changes in one place and every consumer picks it up. AI proposes the definitions and the mappings; your experts confirm them. Every definition is validated with SHACL, a standard TopQuadrant helped author, so the record exports whole. We maintain 300 reusable standards-based models, so "customer" starts from governed structure. Your catalog stays, and it syncs. Our semantic layer use case page shows it.
So keep the metrics layer. Give its nouns an owner. A metrics layer says how much. TQ Data Foundation says what, first, with a name on it.