Google Cloud adds metrics to BigQuery Graph for AI

Google Cloud has introduced measures in BigQuery Graph, now in preview. The update adds metric definitions to graph-based data models in BigQuery.

The change targets organisations building autonomous AI agents that need to analyse business relationships alongside tabular data. Agents working only from raw tables may identify an outcome but miss the chain of dependencies behind it.

BigQuery Graph represents data as connected business entities rather than isolated rows and columns. With the addition of measures, users can define metrics such as sums and averages within a property graph, then query them through standard SQL.

The approach addresses a common data analysis problem: graph structures and key performance indicators are often maintained in separate systems. In those setups, one team may manage supplier or customer relationships in a graph database while another tracks financial or operational metrics in SQL, leaving software agents to combine the two at runtime.

That process can raise costs, slow query execution, and lead to inconsistent KPI calculations. The new feature instead maps existing tables to a property graph in place, without moving the underlying data through a separate extract, transform, and load process.

How it works

The technical issue Google Cloud is addressing stems from standard SQL joins, which can duplicate rows during graph traversals and distort aggregate calculations. In BigQuery Graph, data modelers define a MEASURE directly within the property graph definition language, and the system resolves graph paths before applying aggregation functions.

Users can then query the graph with the GRAPH_EXPAND function alongside the AGG aggregator. The goal is to combine graph traversal for understanding relationships with SQL-based metric calculation, while preserving aggregation integrity.

One company example uses public eCommerce data in BigQuery to create a property graph linking users and orders. In that model, average items per order and total items are defined as measures on the order node, then aggregated by customer city through a graph query.

Agent use cases

Google Cloud framed the release around the growth of agentic AI workloads, in which software agents are expected to act on business questions with less human supervision than simple chat assistants. Those systems can produce inaccurate recommendations if they cannot connect an observed result to the wider web of operational dependencies.

One example described a retailer asking an agent why winter jacket sales fell by 12% in Seattle. In Google's example, a flat-table query could report the decline but fail to trace the sequence from local orders to distribution centres and then to suppliers affected by regional storms, increasing the risk of an irrelevant discount campaign.

Google Cloud is also adding tools in BigQuery Studio to make graph models easier to build and use. These include a visual graph modeler with drag-and-drop controls for defining nodes, edges, and property graphs without writing the full data definition language by hand.

Another addition links BigQuery Graph with Conversational Analytics in BigQuery Studio. Users will be able to ask questions in natural language while the system translates them into GoogleSQL or ISO GQL queries based on the graph's relationship structure.

Google Cloud said the setup is intended to reduce hallucinations by constraining queries to a deterministic map of the data model. It added that graph semantics can reduce the need for users to infer table joins manually before asking a question.

Looker integration

Alongside the BigQuery changes, Google Cloud outlined native integration between BigQuery Graph and Looker analytic models. The aim is to keep business definitions and relationship models in the same data layer rather than spread across separate tools.

Under one method, Looker can point directly to a database-defined BigQuery Graph through sql_analytic_model_name, allowing LookML dimensions and measures to map to graph properties. Under another, users can define the graph schema inside a LookML view with derived_analytic_model, after which Looker generates and runs the SQL data definition statements needed to maintain the graph in BigQuery.

Google Cloud said those options are intended to support analytics team workflows already built around the Looker integrated development environment, Git-based version control, and continuous integration. It said this would help keep core KPIs such as churn rate identical across graph models and reporting systems.

The release reflects a wider push by cloud providers to adapt data platforms for AI systems that do more than retrieve information. Google Cloud is betting that governed metrics and relationship mapping need to sit closer together if businesses want automated agents to explain not just what happened, but why.