Google BigQuery

What it is

BigQuery is Google Cloud’s serverless data warehouse: you load data, run SQL, and pay for storage and compute separately, with no cluster to size. This page covers it as the warehouse behind GTM data, and the part that changed in 2026: Conversational Analytics. Google announced general availability on July 1, 2026. A data team builds a data agent in BigQuery Studio, grounded in chosen tables, views, graphs and UDFs, plus verified queries and written instructions. A business user then asks questions in plain language and gets an answer with a chart, the generated SQL, and citations to the context the agent used. Deep-dive mode handles “why did this metric move” questions with a multi-step plan and a downloadable report.

Agents can be published to Gemini Enterprise (GA since October 14, 2026), to Data Studio, or into your own app through the Conversational Analytics API (GA June 23, 2026). Every query runs under the asking user’s IAM permissions and is logged. Agents are read-only: they cannot run DML, write data or call remote functions. Separately, a Google-managed MCP server at bigquery.googleapis.com/mcp lets Claude and other MCP clients list datasets and run SQL, with results capped at 3,000 rows and a three-minute query timeout by default.

Why it shows up in RevOps and CS stacks

• Google’s marketing data lands there natively. GA4 export and the Data Transfer Service for Google Ads put ad and web data in BigQuery without a third-party pipeline. Add CRM and billing through Fivetran and marketing ops can join spend to pipeline in one place.
• Ops can ask without a BI seat. A RevOps lead asks “which open opportunities over $50,000 come from accounts whose product usage dropped this quarter” and gets SQL they can hand to an analyst for review. No Looker or Tableau license is needed to ask.
• The output feeds the rest of the stack. Segments built in BigQuery sync back to Salesforce or HubSpot through Hightouch. BI tools such as Looker, Hex and Sigma query the same tables.

Pricing reality

Compute is billed one of two ways. On-demand charges $6.25 per TiB scanned, with the first 1 TiB per month free. Capacity pricing charges per slot-hour by edition: $0.04 Standard, $0.06 Enterprise, $0.10 Enterprise Plus, pay as you go, with 1- and 3-year commitments taking Enterprise down to $0.048. Active logical storage costs about $0.023 per GiB per month in US regions, and the first 10 GiB is free.

The agent bills on top of that. Google’s Data Cloud agent pricing charges $3 per million input tokens and $20 per million output tokens for the Conversational Analytics Agent. All Data Cloud agents are in a free trial through December 31, 2026, and the token pricing starts after that. The SQL each answer runs bills as normal BigQuery compute.

Our estimate, at 30,000 input and 3,000 output tokens per question, is about $0.15 in agent tokens. A 40-person ops team asking 1,000 questions per working day lands near $3,150 a month in tokens from January 2027, before query compute. A team on on-demand billing whose questions scan 20 GB each adds roughly $0.12 per question. Measure your own token counts during the free trial instead of budgeting from ours.

Best for

RevOps, marketing ops and CS ops leaders at companies already on Google Cloud, or whose GA4 and Google Ads data already lands in BigQuery, who want 20–200 business users asking pipeline, spend and account-health questions under existing IAM, without buying each of them a BI seat.

Skip it if your warehouse is already Snowflake or Databricks: moving the data to get the agent is rarely worth it. Skip it if users need to write numbers back, such as forecast or quota changes. Data agents are read-only, so that job needs a writeback tool like Sigma.

Versus the alternatives

Snowflake is the top independent warehouse by share, and its agent layer, CoWork, does the same job. Pick Snowflake when you need multi-cloud on AWS or Azure. Pick BigQuery when you are on Google Cloud or most of your marketing data comes from Google products.

Databricks is the other leader. Its business-user agent is Genie, billed in DBUs. Pick it when the data team already runs a Databricks lakehouse for ML. BigQuery can query Databricks Unity Catalog tables as a lakehouse source, but the two are not a mix-and-match pair.

ClickHouse is the fastest-growing entrant. It raised $400 million at a $15 billion valuation in January 2026, per Bloomberg. Pick it for sub-second product analytics on event data. It has no business-user agent comparable to Conversational Analytics.

If none fit because nobody agrees what “pipeline” means, fix the model first. See data warehouse vs CDP and reverse ETL.

Watch-outs

• The free trial hides the agent’s real cost. Token billing starts January 1, 2027, and a pilot measured before then shows only query compute. Guard: during the trial, log input and output tokens per question from a representative week. Multiply by $3 and $20 per million, and get that number approved before rollout.
• Cost caps only work on on-demand billing. The big_query_max_billed_bytes limit and the per-user daily query quotas do not apply to slot reservations, and per-agent quotas and labels are not supported. Guard: run agents in a dedicated project on on-demand billing. Set a per-query byte cap and a “Query usage per day per user” quota there.
• Answers are only as good as the grounding. Raw Fivetran tables with cryptic column names produce confident wrong SQL. Guard: point agents at curated views, not raw tables. Add verified queries for your 20–30 core metrics, and rerun a fixed set of known questions monthly.
• The MCP server can write. Its execute_sql tool accepts DML and DDL, unlike execute_sql_readonly. Guard: give MCP-connected identities only BigQuery Data Viewer and BigQuery Job User on curated datasets, so writes fail at the permission layer.