Guide · See the gold layer

See the gold layer.

Raw source data is refined into clean, business-ready context by a dbt medallion pipeline. Aurelia’s applications, semantic tools and AI clients build on the top layer — gold.

Bronze → silver → gold

After a source lands (see Connect your sources), dbt models refine it through three layers, each building on the one before:

LayerWhat it holdsPurpose
BronzeRaw source data, as landedA faithful copy of what each connector replicated — nothing dropped.
SilverTyped, conformed, de-duplicatedConsistent types, keys and grain across sources; duplicates and noise removed.
GoldDecision-ready martsGoverned business context shaped for applications, tools and agents.

The gold marts

The gold marts live in the workspace’s isolated BigQuery gold dataset. Each one is reusable infrastructure: workspace views inspect it, while Aurelia Auto and MCP clients access it through the governed semantic layer.

MartContents
aurelia_gold_<tenant>.dailyDaily channel performance — spend, reported revenue, orders and new customers.
aurelia_gold_<tenant>.salesSales by region and category — revenue, contribution margin and orders.
aurelia_gold_<tenant>.channelChannel incrementality — MMM incremental revenue and each channel’s modelled share.
aurelia_gold_<tenant>.cohortAcquisition cohorts and predicted lifetime value.
aurelia_gold_<tenant>.dim_metaDataset metadata — date bounds and filter members. Not governed.

Who reads gold

These applications and agents build on the gold layer, so every interface works from the same governed context:

  • Profitability — revenue, contribution and incrementality.
  • Customer LTV — acquisition economics, value and payback.
  • Total commerce — owned and marketplace demand in one view.
  • Customer 360 — unified identities, lifecycle segments and source coverage.
  • Aurelia Auto — reads the same channel-performance marts through the semantic layer to explain results and draft budget moves.

How it’s built & isolated

In production, Airbyte lands raw data and dbt builds each layer in separate tenant-isolated BigQuery datasets. Dataset names are derived on the server from the authenticated workspace; clients cannot choose another tenant’s dataset.

Until a source has synced and the dbt build has completed, a workspace has no gold data yet — applications show an empty state, and a demo workspace uses representative sample context instead.