AI & Gemini in Looker

Every BI vendor shipped a chat box in the last two years. Most of them answer questions by asking a large language model to write SQL against raw tables, and the results are exactly what you would expect: plausible numbers, wrong joins, a "revenue" figure that quietly excludes refunds. Looker's advantage in 2026 is that Gemini does not have to guess. It answers through the LookML semantic layer, so "revenue by region last quarter" resolves to the same orders.total_revenue measure, the same join path and the same access filters your dashboards already use.

That only works if the semantic layer is good. A model full of unlabeled fields, duplicate measures and explores nobody has curated produces confident, wrong answers just as fast as a raw-SQL bot does. Preparing the model is the real work of an AI rollout, and it is work we have been doing for years under a different name.

What we do

Enable and tune Gemini in Looker. We turn on the Gemini in Looker features that fit your edition and licensing, configure the Google Cloud project and Gemini Data Analytics settings, and decide, explore by explore, what should be exposed to natural-language querying and what should stay dashboard-only.

Conversational Analytics rollout. We pilot Conversational Analytics with one business team on one or two curated explores, measure answer accuracy against known-good dashboards, fix the model where answers diverge, and then widen the rollout. Where you need chat-with-your-data inside your own product, we build on the Conversational Analytics API and a data agent grounded in your LookML.

Model curation for natural-language querying. The same things that make a model good for humans make it good for Gemini, only more so:

• clear label and description on every exposed field, written as a business user would phrase the question
• hidden: yes on implementation fields so the model cannot pick a surrogate key when the user meant a name
• one canonical measure per business metric, with the alternates hidden or removed
• curated explores with sensible defaults, always_filter where the data is meaningless without a date range, and access_filter so row-level security applies to AI answers exactly as it does to dashboards

Evaluation of answer accuracy. We build a question set with your team (the thirty questions people actually ask), record the expected answers from trusted dashboards, and run them against the assistant before and after each model change. You get a pass rate you can put in front of leadership, not a demo.

Engagement: two-week AI-readiness assessment

A fixed-scope starting point for teams who want to know where they stand before committing to a rollout:

• Week 1: model inventory and scoring (labels, descriptions, hidden fields, duplicate measures, explore sprawl), permissions review, and a Gemini in Looker feature and licensing check for your instance.
• Week 2: pilot one curated explore with Conversational Analytics, run the evaluation question set, and deliver a written readiness report with a prioritized remediation list and a rollout plan.

The output is a document your own team can execute, or that we can execute with you.

Why Vistelio

We are a Looker-only consultancy. Our senior, US-based developers have built and rescued LookML models across SaaS, healthcare, events and agriculture, and we bring that modeling discipline to AI rollouts rather than treating them as a separate product. Vistelio is an independent consultancy and is not affiliated with or endorsed by Google.

Ready to find out whether your model is AI-ready? Contact us to schedule an assessment.