A Power BI semantic model is the layer that decides what a number is allowed to mean.
It is not the dashboard. It is not the .pbix file you email. It is the tables, the relationships, and the measures every report is supposed to read — so “Revenue” is one thing, not five.
If two pages disagree, the model is usually the argument, not the chart.
What it is, in one sentence
One structured set of tables and relationships, plus a measure table, that every visual is allowed to use.
Star schema is the usual shape: fact tables (invoices, tickets, hours) and dimension tables (date, customer, product, region). Reports sit on top. They do not each invent their own copy of the business.
What it is not
- A pretty report page
- A folder of Excel files with the same column names
- A calculated column pasted into three .pbix files
- A dataset that only the original author understands
If the only way to change a KPI is to hunt through visuals, you do not have a model. You have a drawing.
Why leadership should care
A semantic model is how Monday’s report matches Friday’s.
Without it, each team builds a local version of the truth. Filters drift. Grain drifts. Someone adds tax in one file and not the other. The meeting then spends twenty minutes reconciling, and the decision waits.
With it, the argument moves to the decision. The number is already named.
The three pieces that have to exist
1. Grain
One row equals what? One invoice. One day. One employee-week. If that sentence is missing, every measure on top is a guess.
2. Relationships
How the fact table finds the date, the customer, the region. Bidirectional filters and many-to-many joins are how a total quietly doubles. Relationships should be boring on purpose.
3. Measures
DAX lives in one place. Visuals consume it. They do not rewrite “margin” in a card title. If five people calculate the same KPI five ways, the model is not finished.
Desktop, Service, and the file on a laptop
Power BI Desktop is where the model is built. Power BI Service is where it is published, refreshed, and shared. A .pbix on a laptop after the author leaves is not a model anyone else can trust.
Row-level security, workspace roles, and scheduled refresh belong to that published layer. They are not extras. They are how the same number survives more than one person.
When you already have ten reports
Do not start by drawing a new page. Name the layer that is lying: source, relationships, or DAX. Then fix that layer.
Existing files are starting points. The finished view is one model, a short dictionary of measures, and one to three pages a meeting can open without a walkthrough.
That is the core of PeekGraph’s Power BI consulting: semantic model, DAX, dashboards, and refresh — scoped on a call, not billed from a page.