Gemini can now query your BigQuery and Spanner data in plain English

Google has put Data Cloud connectors into Gemini for business, now in preview. Staff can ask questions in plain language and get answers from live BigQuery, Spanner, AlloyDB, Cloud SQL and Looker data, with each query running under the asker’s own credentials.

How a question becomes a permission-scoped query

According to Google, the connectors work together with Knowledge Catalog, its metadata service. The flow has four steps. A user asks a question without naming a table or project. Gemini then searches the catalog for candidate data, weighing signals such as usage patterns. Technical metadata, business glossaries and semantic models map everyday terms to database structures. Google’s example is resolving “cold weather products” to categories such as sweaters, coats and jackets. Finally, Gemini writes SQL and runs it through the connector.

Google says the execution respects the user’s existing row- and column-level permissions. An employee without access to executive salary data or protected healthcare records cannot reach it through a prompt. Gemini also only sees catalog metadata the user is allowed to see.

Preview status differs by source. Google labels BigQuery, the operational databases (Spanner, AlloyDB, Cloud SQL) and Looker as preview. Results can be moved into a Google Doc, a Gmail draft or a Google Slides deck.

Open tables and AWS data included

The BigQuery connector reaches beyond native storage. Google says it can query open table formats including Apache Iceberg, Delta Lake and Parquet, plus datasets in AWS or Azure through BigQuery. One example in the post shows Gemini answering questions across BigQuery and AWS S3 through the AWS Glue federated catalog. Access policies from BigQuery and Knowledge Catalog carry over to those files.

The connectors use federated query mode, so data is queried in place and not copied or indexed into a Gemini data store.

Setup, and when to pick an agent instead

Administrators configure the connectors in the Gemini admin console. They create default read-only connections to BigQuery, operational databases or Looker, and Knowledge Catalog is enabled automatically. Users then enable the connector in Gemini and authorize it on first use. Admins can set a default billing project for BigQuery queries, which Google says helps control query costs. Documentation exists for the BigQuery, Spanner, Cloud SQL and AlloyDB for PostgreSQL connectors.

Google pitches the connectors for exploratory, ad hoc questions. For repeatable reporting it points to Conversational Analytics agents, which data teams build and certify for a defined slice of data. These can be published in Gemini or exposed through an API. Gemini also surfaces relevant published agents, so a user can choose between a free-form query and a scoped agent.

Google also contrasts the connectors with self-managed Model Context Protocol servers. It says the first-party route brings prebuilt skills for each data engine, such as SQL and AI function generation, and managed authentication that removes the need to run MCP middleware or configure OAuth clients for Google Cloud services.

Rivals already sell governed natural-language SQL

Google is not alone here. Snowflake’s documentation describes Cortex Analyst as a REST API that lets business users ask questions in natural language without writing SQL. Third-party write-ups say it depends on semantic models that teams define up front, and that Databricks’ Genie runs questions over Unity Catalog tables on a SQL warehouse under the user’s identity. Those comparisons come from vendor and consultancy blogs, and one reports an accuracy range for Cortex Analyst that its author could not tie to a primary source, so treat them as context only.

The pattern is similar across the three: a semantic layer supplies business meaning and the database enforces who may see what. Google’s distinguishing claim, based on its own description, is breadth. One connector layer is meant to span Google’s own databases, Looker and open-format or multicloud data in AWS and Azure, inside the same Gemini window where people already write documents and email.

Gemini’s wider push into everyday work shows in recent product moves, such as turning a photo of a sketch into a working app.

What the announcement leaves unsaid

The post gives no accuracy figures, latency numbers or benchmark against other tools, so claims of better SQL quality and lower token use are Google’s own and have not been independently checked. It does not state pricing, regional availability, which Gemini plans include the connectors, or when general availability might arrive. It also does not say how well the catalog performs on messy schemas where metadata is sparse, which is the failure mode Google itself says earlier generative AI tools struggled with. Organizations that want to try the preview will need an administrator to enable it, and the quality of existing metadata and glossaries will likely decide how useful the answers are.