Updated November 16, 2026. Genie Ontology is now generally available: the Databricks release notes list it on November 9, 2026 for AWS, Azure, and Google Cloud. Pages became generally available on November 11, and domains had already done so on October 8. The ontology also has its own documentation page now. It calls the two halves modeled knowledge and learned knowledge, and it says Genie might ask you to clarify a conflict it cannot resolve. We updated the status labels, the conflict stage, the list of gaps, and the sources to match. External knowledge sources are still in Beta.

Databricks defines the Genie Ontology the same way on all three clouds: “the unified context layer that gives AI agents a business-aware map of your organization.” Accurate, and almost useless if you are deciding whether it matters to your company. Here is the plain version.

When someone asks a question of your company data in ordinary English, the system has to work out what the words mean. Which table holds revenue. Whether “active user” counts a person once or once per device. Which of the four dashboards named “Sales” is the one people trust. The Genie Ontology is where those answers live. It is the difference between a system that can read your data and one that understands your business.

The part worth a leader’s attention is not the technology. Half of this layer is defined on purpose. The other half assembles itself, from whatever your teams have already built. Databricks calls the first half modeled knowledge and the second learned knowledge. Click through the figure below to see both halves and what happens when they disagree.

Inside the Genie Ontology Four stages, from what your team defines to the answer a person receives. Every element below is named in the Databricks documentation, except where a note cites a blog post.
Stage 1 · Modeled knowledge

The half somebody has to decide

Databricks calls this modeled knowledge: context that you “define, govern, and certify” through Unity Catalog semantics. The ontology page gives four tools as examples. The two that carried preview labels when we first published are both generally available now.

Metric viewsReusable definitions of a business KPI, governed with normal Unity Catalog permissions. The docs page carries no release-status label.
DomainsGA Groups data by business purpose, so people browse by what something is for rather than by where it is stored. Generally available since October 8, 2026. The REST API for managing domains is still Beta.
PagesGA Governed definitions of a business concept, written for both people and the system to treat as authoritative. Generally available since November 11, 2026.
CertificationMarks an asset as trusted or outdated. The docs now name it among the signals Genie weighs when sources conflict. No release-status label.
This is the half you can direct, and the only half you can hold anyone accountable for. Each of these tools still needs a person to decide. Unity Catalog semantics also gained a fifth tool on November 6: wikis, for documenting a single table, which are in Beta.
Stage 2 · Learned knowledge

The half that assembles itself

Databricks describes this as a map of snippets that Genie “automatically extracts and maintains” from your existing assets and usage. The documented sources are metric views, dashboards, SQL queries, and Genie Agents. Each piece is called a snippet, and the docs name three kinds.

Metric definitionsThe documented example: an “active user” is a distinct user, deduplicated across all platforms.
Authoritative sourcesThe documented example: revenue questions should be answered using the curated Finance Genie Agent.
Business rulesThe documented example: a “qualified lead” only counts after a demo is booked.
Where they come fromMetric views, dashboards, SQL queries, and Genie Agents, including queries that outside tools such as dbt run on Databricks. Definitions from BI tools such as Power BI can be added as external knowledge sources, which are in Beta and stay off until an account admin turns them on.
Read that list of examples again. Those are the definitions that currently live in analysts’ heads and in footnotes on slides. This is where a company’s quiet disagreements about what counts as revenue finally surface.
Stage 3 · The authority score

What happens when two definitions disagree

Each snippet carries an authority score, and Genie weighs the competing sources at the moment the question is asked. The documentation names three inputs to the score.

Where it came fromThe asset it was generated from
How often it is usedUsage of the snippet
How fresh it isRecency of the snippet

Genie then ranks the most relevant snippets and resolves conflicts by assessing the authority of each source. The docs give certification, popularity, and author as example signals, and they state the precedence rule two ways. One section says modeled knowledge takes precedence when the two conflict. The conflicts section says modeled knowledge, certified assets, and curated Genie Agents “generally take precedence” over learned knowledge on the same topic. Because of that word, generally, we still read this as prioritization rather than a guaranteed override. And if Genie cannot resolve a conflict, the docs say “it might ask you for clarification.”

Worth knowing: three Databricks sources describe the ranking, and they do not list the same things. The June launch post compared it to PageRank and gave five factors. The November 11 general-availability post names it OntoRank and lists source authority, certification, the people behind the knowledge, popularity, and recency. The docs name three inputs to the score and three example signals for conflicts. None of them publishes the weights or the algorithm itself.
Stage 4 · The answer

Permissions still decide what anyone sees

Snippets are gated by Unity Catalog permissions. Databricks states that Genie answers “using only the sources you are permitted to see.” For assets inside Databricks, the ontology changes how well a question is understood, not what anyone can reach. The docs note one exception: snippets created from external knowledge sources, the Beta feature, are available to all users in the account.

QuestionAsked in ordinary English
Ranked contextOnly snippets this person may see
AnswerWith citations, and the snippets it used one click away
The same ontology now serves more than chat. The docs say Genie One and Genie Code both search it, and agents on other platforms can reach it through the Genie One MCP server with the same Unity Catalog permissions. On speed, the current ontology page makes no claim. The November 11 post says an agent handed the right definitions can skip open-ended exploration of the data.

What this actually changes for a business

The honest summary is that the Genie Ontology moves an argument, rather than settling it. Companies have always disagreed about what a metric means. Until now that disagreement was survivable, because each team ran its own report and the numbers only collided in a meeting once a quarter.

Put a system in front of everyone that answers in plain English, and the collision happens every day, in public, in front of whoever asked. So the work is not technical. Someone has to decide what “revenue” means and write it down where the system can read it. That is a leadership task wearing a data-tooling costume.

Four things the documentation does not promise

We read the primary sources rather than the launch material, and the gaps are as informative as the features.

  • General availability covers the core, not everything around it. The release notes list Genie Ontology as generally available on November 9, 2026, on AWS, Azure, and Google Cloud. External knowledge sources are still Beta. Ontology Manager, the interface for curating the ontology, is not in the docs: the November 11 Databricks blog post calls it “coming soon in Beta.” On Genie Search the sources disagree. The AWS and Azure release notes list it as generally available on November 9, and the Google Cloud notes do not list it at all. The November 11 press release says Genie Search, Ontology Manager, and the always-on Genie are “coming soon to preview and beta programs.” We go by the docs, and we name the source each time.
  • You cannot edit an individual learned snippet. The docs now say so directly. They do describe how to see them: click Genie Ontology under a response to open the snippets it used, each with its source asset, author, domains, and lineage. There is still no documented way to approve, downrank, or delete one. You steer it indirectly, by updating the source asset, by restricting access to stale assets, or by defining the concept as modeled knowledge.
  • There is no documented off switch. The ontology was turned on by default on August 6, 2026, while it was still in preview, and the page that documents it today describes no setting to turn it off. The documented controls are narrower. Unity Catalog permissions keep an asset’s snippets away from people who cannot see the asset, and an account admin can turn off the external knowledge sources preview.
  • The accuracy numbers are the vendor’s own. The June launch post footnoted its benchmark as an internal test of 28 questions. The November 11 post adds a second benchmark, built with Turing, in which Genie passed 91% of task runs. Databricks published it, the rival agents are anonymized, and the post says the results reflect a deployment where that knowledge had already built up. We have not found an independent measurement of the ontology’s contribution. Treat the figures as a vendor claim, because that is what they are.

What we would do about it this quarter

Nothing here requires a project. It requires a decision and an owner.

  1. Pick the five metrics your leadership actually argues about. Revenue, active user, qualified lead, churn, and whatever your industry adds. These are the ones worth defining first.
  2. Give each one a written definition and a name. Pages and metric views exist for this, and Pages left Beta on November 11. The tool matters less than the fact that a person decided.
  3. Name an owner per definition. Not a committee. When the definition needs to change, someone has to be allowed to change it.
  4. Assume it is already running. The learned half has been building itself from your dashboards and queries since August. The question is not whether to start, it is whether anyone is watching what it concluded. You can check now: open the snippets under an answer and read what it learned.

Here is the uncomfortable version. The system is already forming a view of what your business terms mean, assembled from whatever your teams happened to build. If nobody writes the real definitions down, that learned view becomes the definition by default.

Where these facts come from

Every claim above traces to Databricks documentation or an official post. When we first published in August, the ontology had no documentation page of its own, only a section inside the “Chat in Genie One” page. It has one now, last updated November 9, 2026, and most of the facts above sit there. The definition and the conflict rules read the same on AWS, Azure, and Google Cloud. The links below go to the AWS version.

One correction to our own earlier reporting, and then a correction to the correction. In our July newsletter we wrote that steward certification feeds into the Genie Ontology. In August we said the documentation did not support that as written. It does now: the ontology page lists certification among the modeled knowledge features, and names it as a signal Genie weighs when sources conflict.