The Ontology Market: A Guide to Five Very Different Things
Key takeaways
- The ontology market is in peak confusion, and that is a sign of investment rather than a problem.
- Five distinct artifacts now carry the name, and each answers a different question.
- Most buying confusion comes from comparing products that are in different categories.
- Two questions cut through it: does the model survive leaving the platform, and what else does it connect to?
This guide maps the ontology market, which is in peak confusion. That is a good thing.
Ontology is having its moment because the previous term lost its meaning. Semantic layer once meant meaning that spanned the enterprise. It now means BI and metrics, and the thing it used to describe needed a new name. At the same time, surging investment in data infrastructure has produced several genuinely different artifacts, all reaching for the same word.
So "ontology" now covers at least five things. They solve different problems. They are built by different kinds of company. Most of the confusion in a buying process comes from comparing two of them as though they were alternatives. Here are the five, as each vendor describes them in October 2026. This market is moving fast enough that the categories will outlast the placements.
| Category | What it models | The question it answers | Where it lives | What survives if you leave |
|---|---|---|---|---|
| Metrics semantic layer | Dimensions, measures, joins | How is this number calculated | Over the warehouse | The definitions, if the syntax is open |
| Catalog glossary with a graph | Terms, owners, lineage, access | Where is this data and can I trust it | In the catalog | Export of terms, usually without the graph |
| Learned context graph | Inferred term meanings, ranked | What do people here probably mean | Across apps and chat | Little. The value is the inference |
| Entity and object-action model | Objects, links, actions | What exists and what may be done to it | In the platform | Varies by vendor. Ask whether the model exports to open standards |
| Ontology of record | Classes, properties, constraints | What is true, who agreed it | Anywhere. It is a standards file | The model, its rules and its identifiers |
1. What is a metrics semantic layer?
Revenue gets one definition. These model dimensions, measures and joins over the warehouse so that a dashboard, a report and a question asked in natural language all return the same number.
They say how to calculate something. They do not say what a thing is, and they are not trying to. That distinction sounds pedantic until an agent needs to know whether two records are the same customer, which is a question no metrics layer is built to answer.
Who ships them: dbt Labs MetricFlow, open source and the most portable of the group. Looker's LookML, now grounding Google's conversational analytics. Snowflake's semantic views, grounding Cortex agents.
When it is the right choice: your problem is that three teams compute the same metric differently, and everything you need already lives in one warehouse.
2. What is a catalog glossary with a graph on top?
This one drives the most confusion, because the artifact is more than a decade older than the word now attached to it. A business glossary with terms, owners, lineage and access control, with a graph layer added so the relationships between assets are navigable.
The governance value is real. Someone can find a table and know whether to trust it, who owns it and where it came from. What it describes is the catalog, and the data estate, more than the business behind them.
Who ships them: Collibra, running semantic agents over a governed glossary. Atlan, a context lakehouse serving over MCP. Alation, with semantic model mastering and some of the clearest published writing on exactly these distinctions. Their Ontologies product entered early access in September 2026, so what it documents will have moved by the time you read this.
When it is the right choice: discovery and trust are the bottleneck. People cannot find data, or cannot tell which copy is the good one.
3. What is a learned context graph?
Trends and inferences, stored as reusable context. The system reads your tables, your dashboards and your chat, works out what your terms most likely mean, and ranks the competing versions it finds.
This is the most dynamic of the five and the closest thing to an active model that exists. It is also the category where the output is a well-informed guess, which is the right trade for some problems and the wrong one for others. An inferred definition can tell you how a term is currently used. It cannot tell you what anyone agreed, which is the distinction the whole ontology management discipline turns on.
Who ships them: Databricks Genie Ontology, It extracts definitions from across the stack, including 50+ connected apps and data systems. It then ranks them by an approach Databricks describes as "similar to PageRank", weighing where a definition came from, how often people rely on it and how fresh it is. Glean, an Enterprise Graph across 275+ out-of-the-box connectors.
When it is the right choice: you need coverage fast across a sprawling estate, and approximate meaning is more useful to you than no meaning.
4. What is an entity or object-action model?
What exists, and what an agent may do about it. Verbs as well as nouns: an order, its shipment, the action "reroute", and who is permitted to trigger it.
This is the most operationally useful of the four preceding categories, because it closes the loop between describing something and acting on it. The model is expressed in the platform's own language, which is what makes it immediately executable there and what makes it the platform's.
Who ships them: Palantir Foundry, Its primitives are object type, link type and action type. Its documentation defines them as the schema definitions of an entity, a relationship between two entities, and a set of edits to them. Ontology-as-code exists there too, in beta. Microsoft Fabric IQ, with entity types carrying properties, relationships, metrics and rules, where Microsoft defines a rule as "a natural-language statement of business logic linked to ontology concepts."
A useful note for anyone assuming this category is a one-way door. It is not uniformly so. Microsoft documents that Fabric IQ can import RDF, Turtle and OWL definitions and export in RDF or Turtle. Palantir's documentation contains none of those standards. Both are object-action models, and their exits are not the same. Ask about the exit. Do not assume it.
When it is the right choice. You have committed to the platform. The operational use case is inside it. Speed to a working agent matters more than portability. More on that trade in what Palantir means by ontology.
5. What is an ontology of record?
Machine-readable and authoritative beyond any one tool. This is the former promise of the semantic layer, which is why the term needed replacing.
The category is defined by four properties. It is built on open W3C standards, so it is composable and extensible. It carries global identifiers that survive the system they were created in. Its rules are expressed as code, in SHACL, so a validator can check them. And an agent's plan can be validated against it before the plan runs.
Who ships them: Protégé, TopQuadrant, and other semantic platforms.
When it is the right choice. The meaning has to outlive the tool. More than one system depends on it. Or somebody has to be able to say who agreed a definition, and when. The practical argument for the category is in what an ontology is. The argument against doing it at all, taken seriously, is in do you actually need an ontology.
Which questions is this market still sorting out?
Both are worth putting to a vendor in any of the five categories, and the answers separate them faster than a feature list does.
- Does your ontology need to live outside the platform that bundles it? Governance that was optional becomes the tier you are on. Databricks is retiring its Standard tier and will automatically upgrade any remaining workspace to Premium on 1 October 2026; Unity Catalog governance lives in Premium. Products also stop. Grafo, a knowledge graph design tool, states on its own homepage: "Grafo was retired on March 13, 2026." Bundling is not the question. The question is whether this particular model is one you need to keep.
- What else does it connect to? Direct connection to the data is table stakes. Master data, rules and policy, and unstructured sources are where it gets decided. A model that knows what a customer is, but not which record survived mastering or which policy governs the action, grounds half a decision.
The confusion in this market is a market sorting itself out, and it resolves the moment you can name which of the five a vendor is selling. Most of the time they are not competing with each other at all.
So what should I take away?
Five distinct artifacts now carry the name. Two questions cut through it. Does the model survive leaving the platform? What else does it connect to?
Answer those two and the category sorts itself, and so does the shortlist. TopQuadrant sits in the fifth category, the ontology of record. It is the only one of the five that a standards body defines instead of a vendor.
