Google Cloud links Gemini Enterprise logs to BigQuery

Google Cloud has outlined a way for organisations to analyse and govern Gemini Enterprise deployments through BigQuery by moving application telemetry and audit data into Google's data warehouse.

The setup is aimed at administrators overseeing large Gemini Enterprise roll-outs, where prompt logs, model responses, user activity and audit records can create a heavy monitoring burden. It combines continuous log routing for detailed records with batch exports for aggregate usage metrics.

Under this model, telemetry is split across five BigQuery tables: user prompts, model outputs, user activity records, cloud audit activity logs and cloud audit data access logs. A separate batch export table holds aggregated seat and engagement metrics.

That gives IT, data and security teams a single place to query how staff are using the software, which departments are building custom agents, how often NotebookLM is used and what data sources are accessed during grounded searches. The same data can support compliance reviews, including checks on file access paths and administrative changes.

Logging design

The proposed ingestion pipeline uses Cloud Logging log router sinks to stream conversational data into BigQuery. A separate audit logging path captures administrative actions and data access activity tied to Discovery Engine services.

Admin Activity logs are enabled by default, while Data Access logs must be turned on separately in Google Cloud IAM settings for the Discovery Engine API. That distinction matters because user-level read and write activity during chats is not recorded unless those settings are enabled.

Prompt and response logging must also be enabled in the Gemini Enterprise admin console before conversational telemetry can be collected. Without that step, administrators would not have the verbatim inputs and outputs needed for detailed reviews.

Analysis tools

Once the data is in BigQuery, Google positions its built-in AI features as the main way to interrogate those records. It highlighted Conversational Analytics in BigQuery Studio, which generates and runs SQL from natural language prompts against the underlying schema and business metadata.

The aim is to reduce the complexity of querying nested JSON log structures, which often require specialist SQL knowledge. This can help administrators who want quick answers on adoption patterns, safety incidents or usage comparisons between tools such as NotebookLM, Deep Research and custom agents.

BigQuery can also generate table and column descriptions, profile datasets and suggest join paths across multiple tables. That allows administrators to inspect fields such as user identities, finish reasons and grounding content, then build a clearer map of how usage data, model outputs and audit events relate to each other.

Dataset-level insights can also produce an interactive relationship graph and suggest cross-table queries. In practice, that could support work such as linking user activity with model responses to measure task complexity or tracing anomalous behaviour across multiple logs.

Governance focus

A central part of the proposal is governance rather than simple usage reporting. Organisations can use BigQuery to investigate security alerts, including cases where content filters block prompts, by locating the exact historical text linked to an alert.

Google also pointed to checks on grounding queries across file repositories and corporate directories, which could help companies review potential exposure of internal information. The approach is designed to supplement, rather than replace, the precomputed dashboards already available in Gemini Enterprise for day-to-day adoption and engagement tracking.

For senior managers, BigQuery data can be connected to Data Studio dashboards to present metrics such as user adoption by department, the ratio of custom agents to employee headcount, connector usage across services including SharePoint, Google Drive and Gmail, and trends in Model Armour blocks.

Google also suggested customers could combine Gemini Enterprise logs with HR or line-of-business data to estimate time savings and build internal reporting for executives. That would move analysis beyond product usage into broader organisational measurement, though the quality of those results would depend on how companies join and interpret their internal datasets.

The overall goal is to give lean administrative teams a more practical way to monitor large-scale AI deployments without relying on bespoke software. Google described BigQuery as the foundation for deep-dive analysis, compliance audits and internal dashboards built on Gemini Enterprise telemetry.