Scoring model
How we score data catalogs for AI agents
Short answer
We score nine platforms from 0 to 10 on nine weighted criteria, using public vendor pages. Totals are computed in code from the weights on this page.
By The Lineage Desk · Best Data Catalogs · Published
Desk research from public vendor pages. Editorial assessment.
What question does the ranking answer?
Which platform gets enterprise data ready for AI agents fastest and most reliably. Ready means an agent can find the right asset, understand what it means, know whether it can be trusted and stay inside the access rules, without the data team hand-writing that context for every question.
Which criteria and weights do we use?
| Criterion | Weight | Question |
|---|---|---|
| Agent access to context | 15% | How do AI agents get the context: an MCP server, context APIs, SDKs, named agent clients? |
| Cross-stack lineage | 13% | Does lineage run from the warehouse through transformation and BI to the agent, at column level, as the vendor documents it? |
| Automated context build | 13% | How much of the context is mapped, documented and kept current automatically, and how soon does the vendor say first value arrives? |
| Governance agents can use | 13% | Can certification, sensitivity and access rules travel with the context, scoped to the agent or the person it acts for? |
| Business definitions and metrics | 12% | Does it hold or extract metric and business definitions that agents can ask for by name? |
| Answer quality checks | 8% | Does the vendor describe a way to test agent answers, catch context drift or trace what agents did? |
| Integration breadth | 10% | How many sources and tools does it connect to, as published, and are there plan limits? |
| Pricing transparency | 8% | Are prices, plan limits or a free tier published? |
| Deployment options | 8% | Can you run it as SaaS, self-hosted, open source or inside your own warehouse? |
Why these weights?
Agent access carries the most weight (15%) because context an agent cannot reach does not help it. Lineage, automation and governance (13% each) decide whether that context is complete, current and safe to use. Business definitions (12%) decide whether an agent's numbers match the dashboard. Integration breadth (10%) matters because few enterprise stacks come from one vendor. Answer checks, pricing transparency and deployment options (8% each) affect reliability over time, budget and fit with security policy, but none of them makes data agent-ready by itself. No single criterion carries more than 15%.
What does each score mean?
| Score | Meaning |
|---|---|
| 9 to 10 | Documented in detail on the vendor's pages and generally available. |
| 7 to 8 | Documented, with a gap: a missing detail, a narrower scope or a feature in early access. |
| 5 to 6 | Partly documented, or described without enough detail to rely on. |
| 3 to 4 | Weak, early access only, or not stated where the vendor would be expected to state it. |
| 0 to 2 | Not described on the pages we read, or outside what the product sets out to do. |
Scores are whole numbers with a one-line reason and a source link on every profile. The total is the sum of score times weight divided by 100, shown to two decimals. Ranks use the exact total; equal totals share a rank.
What counts as evidence?
- Vendor pages: homepages, product pages, pricing pages, documentation, press releases and marketplace listings.
- A few minor facts come from search-result summaries of vendor pages that would not open; they are marked as such next to the source.
- Vendor figures, customer results and analyst placements are linked to their sources.
- Review-site ratings are not used or shown, because we did not read them on the rating sites themselves.
- Where a fact is missing we write Not published or Not described on the pages we read, and the score reflects that.
What does the context path show?
Every platform gets a context path: five hops that context has to travel before an agent can use it. In order: 1. Warehouse and lakehouse; 2. Transformation (dbt, ETL); 3. BI and semantic logic; 4. Governance and access rules; 5. Delivery to agents (MCP or API). A filled node means the vendor's own pages describe that hop. It is a record of documentation. Here is Euno's path as an example.
- Documented on the vendor's pages
- Not documented on the pages we read
1Warehouse and lakehouse: Documented
Live lineage from the warehouse; Snowflake and BigQuery appear in the AlphaSense stack.
2Transformation (dbt, ETL): Documented
dbt models in the lineage graph, plus dbt and Looker sync.
Source: Euno: dbt and Looker use case
3BI and semantic logic: Documented
Definitions inside BI workbooks reach agents through lineage and SQL analysis.
Source: Euno: Context for AI Agents
4Governance and access rules: Documented
Governance rules become live labels; context is scoped by persona.
5Delivery to agents (MCP or API): Documented
MCP server and Context API.
Source: Euno: Context for AI Agents
Who is not scored, and why?
- Stardog: Our reads did not cover lineage, metadata automation or BI coverage, and Stardog's MCP server appears only in a secondary source, so too many criteria would be guesses.
- data.world (ServiceNow Data Catalog): data.world is now sold as ServiceNow Data Catalog with the ServiceNow AI Platform. Its own site could not be opened, and agent delivery, lineage and time to value were not stated on the ServiceNow page we read.
- Modus: Modus is early-stage. Integrations, customers, lineage depth and pricing were not stated in the page text we read, so most criteria could not be scored from facts.
- Jedify: jedify.com could not be opened, so product pages, pricing and agent delivery were not read; MCP is not mentioned in the funding release we read.
Snowflake Horizon and Databricks Unity Catalog are features of the platform you already run, so we describe them and do not score them against products.
How do we handle corrections?
Send a correction through the contact page with the topic Correction to vendor information and a link to the source. We check it against the vendor's page and update the data; totals recompute automatically.
Frequently asked questions
Can a vendor pay to change its score?
No. Scores change only when a published source changes.
Can I use my own weights?
Yes. The weight-your-own page re-ranks the nine platforms with your weights and shows the formula it uses.
Read next
- The ranking
All nine platforms in computed rank order.
- Weight your own
Re-rank the list with your own weights.
- Platform-native options
Snowflake Horizon and Databricks Unity Catalog.
- Reviewed, not scored
Platforms our sources did not cover well enough to score.