Start Gemini in Looker on one or two curated Explores, and open each new business domain only after its data agent answers that domain's critical questions through verified queries. Google's own rollout guidance follows the same order.
When you need this
- Leaders want business users asking data questions in natural language and no longer filing requests.
- Conversational Analytics is on, and early answers picked plausible but wrong fields.
- Data agents are about to be published to Gemini Enterprise or embedded in another application.
Evaluation criteria
Setup: enable deliberately
Which instance settings and permissions turn Conversational Analytics on?
Look forGemini enabled for the instance, from the Google Cloud console on Looker (Google Cloud core). Then the Conversational Analytics toggle on, and access_data, gemini_in_looker, and chat_with_agent granted through a custom role on curated models only.
Setup: curate the Explores
Are the first Explores ready for an agent?
Look forOne or two Explores on clean data, with business labels, descriptions, and synonyms, and with technical fields hidden. Agents cannot use custom fields, so move the important ones into LookML.
Rollout: specialized agents
Does each agent serve one domain?
Look forA Revenue or Operations agent connected to a few closely related Explores, with instructions, default filters, and a glossary. One agent over all company data answers worse.
Governance: verified queries
Which questions must the agent answer exactly right?
Look forVerified queries, generally available since August 28, 2026, for the critical questions in each domain, updated alongside LookML whenever a definition changes.
Review practice: watch feedback
Who reads the feedback and engagement data each week?
Look forThe Conversational Analytics dashboard in System Activity, user ratings, and, where users consent, query review (pre-GA), with fixes made in LookML first and agent instructions second.
Cost control: baseline tokens
What will this usage cost after the promotional period?
Look forToken use by agent and surface, measured now against the instance's monthly allocation, and warehouse compute budgeted for the extra queries. One pilot admin reports about 250,000 input tokens per question on agents built on 80-field Explores, and more on dashboard agents.
Questions to ask vendors
- Which domain has the most repeated questions and the cleanest LookML to start with?
- Who owns each agent's instructions and verified queries once it launches?
- Will agents be published to Gemini Enterprise, and who approves that?
- What share of test questions must an agent answer correctly before a new group gets access?
Red flags
- Every user granted the Gemini role on every model on day one.
- Agent instructions that redefine metrics differently from LookML.
- No verified queries for the questions leadership will ask first.
- Pre-GA features turned on for regulated data.
Products in this piece
Sources
- Official documentation
- 7
- Public source
- 5
- Looker docs: Recommended setup and rollout strategy for Conversational Analytics
- Looker docs: Set up Conversational Analytics in Looker
- Looker docs: Best practices for configuring Conversational Analytics
- Looker docs: Create and manage Explore data agents
- Looker docs: Administer Gemini on your Looker (Google Cloud core) instance
- Looker pricing
- Looker release notes
- Reddit comment on a year to gain buy-in for Looker (Nov 22, 2025)
- Reddit post on rolling Looker out to finance teams that work in Excel (Oct 7, 2026)
- Reddit post on reusing metrics across Explores at scale (Apr 17, 2025)
- G2 review on modeling metrics once in LookML (Nov 15, 2025)
- Google Developer forums: Conversational analytics / gemini token usage data (ruthc and others, Sep 14, 2026)