LLM-ready BI metadata: make your semantic models answer AI search correctly
Search engines and LLMs do not read your DAX. They read names, descriptions, and surrounding docs. If metadata is empty, AI will invent definitions.
What AI systems actually consume
| Source | Used by | Your action |
|---|---|---|
| Table/column descriptions | Copilot for Power BI, Fabric | Write plain English definitions |
| Measure names + descriptions | NLQ, agents | Avoid Measure7; use Net Revenue |
| Report titles & workspace names | Enterprise search | Consistent domain vocabulary |
| External docs (llms.txt, wiki) | LLM crawlers, RAG | Link canonical definitions |
Treat semantic model metadata like public API docs: precise, boring, maintained.
Minimum metadata standard
For every certified model:
- Table description: grain, source system, refresh SLA
- Column description: business meaning, allowed values, PII flag
- Measure description: formula intent in words, not DAX paste
- Synonym line: “Also called: net sales, revenue after returns”
Copilot uses this context. So do internal GPT bots wired to your data catalog.
llms.txt and AI SEO for BI teams
Public-facing BI guidance (blog, internal portal) should mirror model language. If the model says Customer Lifetime Value, your docs should not say CLV metric v2.
Publish a plain-text summary page listing:
- Certified datasets and what questions they answer
- Owner contact and freshness guarantees
- Links to glossary entries
Same pattern as llms.txt for websites — structured, crawlable, no JavaScript required.
Governance hook
Block promote to Prod if top measures lack descriptions. Tabular Editor scripts can export metadata to JSON for CI diff. AI quality follows metadata quality.
Want a metadata sprint on your top three models? Book a 30-minute call.