← CDO Compass
Technology & Innovation

Choosing a Data Catalog

A vendor-neutral guide for CDOs: what to settle first, what to ask, how to test, and the 2026 market at a glance

Most data catalogs are bought after a good demo and judged a year later on whether anyone uses them. Those are two different tests. This article is about passing the second one before you sign.

A data catalog is one of the most visible purchases a data office makes. It is the tool people will point to when they ask what governance has delivered. That makes it tempting to buy the one with the best search bar, the smoothest lineage graph and the most impressive AI features. All of those can matter. None of them decides whether the catalog will be full, trusted and used a year later.

This guide is vendor neutral on purpose. It does not rank products, and it names none of them as better or worse. What it offers instead is a method: what to settle before the first vendor call, which questions separate a real capability from a polished slide, how to run a proof of concept on your own data, and which signs in a demo should make you slow down. I have worked hands-on with open source metadata tooling in my own Labs, and spent most of my career on the governance side of the table. The method below comes from both.

Part I · The Problem

Bought on Features, Judged on Adoption

Vendors compete on features because features are easy to demonstrate. A connector library, a lineage graph and a natural language search can all be shown in thirty minutes. Adoption cannot. Nobody can demo the moment, six months after go live, when a financial analyst looks for the definition of "active client" and either finds a trusted answer or gives up and asks a colleague.

The result is a familiar pattern. The catalog is installed, connectors are switched on and thousands of tables appear overnight. The first weeks look like success. Then people search, find technical names with no descriptions and no owners, and stop coming back. The catalog is not wrong. It is empty in the way that matters: full of metadata, short of meaning.

That is why the most important part of a catalog selection happens before any vendor is invited. If you do not know which people the catalog serves, which questions it must answer for them and who will curate the content, every product will look good in a demo and none will succeed in production.

A catalog full of metadata and short of meaning is the most expensive empty room in the data office.

Bernard Millet, DataEdgePro.net

Part II · The Value

What a Well Run Catalog Gives You

Before choosing a catalog, be clear about what you expect it to deliver. The answer shapes everything that follows, from the criteria you weight to the people you involve. A catalog that is curated and used pays back in at least six ways, and compliance is only one of them.

BenefitWhat it looks like in practiceWho feels it first
Faster discoveryPeople find the right data set, with its definition and owner, in minutes instead of asking around or rebuilding what already exists.Analysts, data scientists, new joiners
Trust and transparencyEvery key data set shows who owns it, where it comes from, what it means and whether it is certified. Debates about "whose number is right" get shorter.Business users, management, finance
Better data qualityQuality checks and their results sit next to the data they describe, and every issue has a visible owner to route it to. Problems are seen at the source, not in the board report.Data owners, stewards, report producers
AI readinessModels, assistants and agents need context: which data is approved, what it means, how sensitive it is and where it came from. A curated catalog is where that context lives.AI and data science teams, risk
Safer changeBefore a column, table or system changes, lineage shows which reports, models and processes depend on it. Fewer broken dashboards after a release.Data engineers, IT, change teams
Compliance evidenceClassified personal and critical data, owners and lineage that can be shown to an auditor or regulator on request.Data protection, compliance, audit

AI readiness deserves a closer look. A model trained or prompted on data nobody understands produces answers nobody can explain. The questions every AI use case raises are metadata questions: is this data approved for this purpose, what does this field really mean, does it contain personal data, how fresh is it, and where did it come from? A catalog that answers them consistently is the foundation of responsible AI, not a side project next to it.

Data quality is the other benefit that tends to be underestimated. A catalog does not fix bad data. What it does is make quality visible and owned. When quality scores appear next to a data set in search results, people choose the reliable source. When every issue has a named owner, it gets fixed upstream rather than patched in every downstream report.

Compliance remains a strong argument, and in Switzerland it is often the one that unlocks budget. Under article 12 of the revised Federal Act on Data Protection, controllers must keep a record of processing activities covering, among other things, the purpose of processing, the categories of data subjects and personal data, the recipients and, where possible, retention periods.[1] For banks, FINMA Circular 2023/1 expanded expectations on identifying and protecting critical data.[2] A catalog can support both. It is not a substitute for either. Ask whether the tool can hold and evidence what those obligations require, not whether the vendor says it is "compliance ready".

Every AI question is a metadata question in disguise: what does this data mean, can we use it, and where did it come from?

Bernard Millet, DataEdgePro.net

Part III · Before the First Call

Settle Three Things Before Talking to Vendors

First, the primary use case. Catalogs are sold as platforms that do everything. In practice, each organisation starts with one or two jobs, and the right choice depends on which. Self service discovery for analysts asks for strong search and business friendly descriptions. Governance and stewardship asks for workflows, ownership and certification. Impact analysis for engineering teams asks for deep, accurate lineage. Data quality management asks for quality signals shown next to the data and clear routing of issues to owners. Regulatory evidence asks for classification, inventories and audit trails. Preparing data for AI asks for all of the above, with a strong API. Pick the first one or two. Write them down.

Second, the first domain. Do not plan to catalog everything. Choose one business domain with a willing owner, a known pain point and data that people already argue about. Finance, client data or a regulatory reporting domain are typical candidates. The first domain is where you will prove value, and it is also what you will use in the proof of concept.

Third, the curation model. Decide who will write descriptions, link glossary terms and certify assets, and how much of their time that will take. If the honest answer is "nobody yet", that is the first problem to solve, and no tool will solve it for you.

Part IV · The Questions

Eight Questions That Separate Capability from Slides

The table below is the core of the evaluation. Each line has a question to ask and the kind of answer that should reassure you. Ask for evidence, not adjectives.

AreaQuestion to askA good answer looks like
ConnectorsWhich of our exact systems and versions do you connect to natively, and what metadata does each connector extract?A list matched to your brief, with clear limits per connector. Not a logo wall.
LineageHow is lineage captured: parsing SQL, reading query logs, from transformation tools, or added by hand? At table or column level?A clear method per source, and an honest view of where lineage will have gaps.
CurationShow us how a steward adds a description, links a glossary term and certifies an asset. How many clicks?A short, simple workflow a business user could follow without training.
GlossaryCan business terms have owners, approval workflows and links to physical columns?Terms that are governed objects, not free text labels.
Sensitive metadataDoes profiling store sample values? Where? Who can see them, and can it be switched off per source?Fine grained control, with sampling configurable or disabled.
HostingWhere is the service hosted and where is metadata stored? Is self hosting possible?A precise region, a documented subprocessor list and a clear answer on options.
OpennessCan we export all metadata, including curation, through an API, in a usable format?A full, documented API and a tested export. Your curation work is your asset.
Cost modelHow does price grow: per user, per connector, per asset, per compute? What happens in year three?A model you can project for your growth plan, with no surprises tied to adoption.

Two lines deserve more attention than they usually get. The first is sensitive metadata. Many catalogs profile data to show value distributions and samples. That is useful, and it also means real values from your systems may be copied into the catalog. For a Swiss bank, that can include client identifying data. For any company, it can include personal data that now sits in a new system with its own access rights. Ask the question early, and make sure your data protection and security colleagues hear the answer.

The second is hosting. If the catalog is a hosted service, it becomes a third party that holds information about your data estate. For Swiss banks and securities firms, the institution remains responsible for the proper performance of what it hands to third parties.[3] That does not rule out hosted catalogs. It does mean your outsourcing, security and data protection reviews belong in the selection timeline, not after it.

Which Criteria Weigh Most for Which Use CaseAuthor's framework · filled circle means decisive
Discovery Stewardship Impact analysis Data quality Regulatory AI readiness Search and descriptions Curation workflow Governed glossary Quality signals and routing Column level lineage Classification and audit trail Open API and export

Read the matrix by column, starting with your primary use case. The filled circles are where a weak answer should rule a product out. The open circles still matter, but they can be traded against price or ease of use. Two patterns stand out. Curation workflow is decisive in five of the six columns, yet it is easy to underweight, because it is the hardest thing to show in a demo. And the AI readiness column is filled almost entirely. That is the point: AI does not need one more feature, it needs everything the other use cases need, done consistently.

Part V · The Proof of Concept

Test on Your Data, with Your People

A demo shows what a product can do on data the vendor prepared. A proof of concept shows what it will do on yours. The differences tend to show up in the same places: connectors that work less fully on your versions, lineage that breaks on your stored procedures, and curation screens that confuse your stewards.

Keep the proof of concept short and narrow. Use the first domain you chose in Part III. Connect its real sources in a non production environment. Then run three scenarios, each tied to one of the five user questions in your brief, and let real users run them, not the vendor's engineers.

What a Good Proof of Concept Tests

  1.   Connect the real sources of one domain yourselves, with vendor support, and record what was extracted and what was missing.
  2.   Trace one known report back to its source columns and compare the catalog's lineage with what your engineers know to be true.
  3.   Ask a business steward to document and certify ten assets on their own, and time it.
  4.   Ask an analyst who was not involved to find the answer to one of the five user questions, and watch where they get stuck.
  5.   Export the curated metadata through the API and check that you could load it somewhere else.
  6.   Ask security and data protection to review what the catalog stored, including any sampled values.

Run the same scenarios on every shortlisted product, with the same people. Score each one on what happened, not on what was promised. If a vendor will not support a proof of concept on your data, that tells you something about how the implementation will go.

Part VI · The Demo

Red Flags, and What to Ask Instead

Demos are designed to impress, and there is nothing wrong with that. The task is to see past the impression. The signals below do not disqualify a product on their own. Each one is a reason to ask a harder question.

What the demo shows

  • A perfect lineage graphClean, complete, end to end, on a sample warehouse built for the demo.
  • AI that writes descriptionsEvery table documented in seconds, with confident business language.
  • A long logo wallHundreds of connectors, shown as icons.
  • Customer stories by nameLarge brands, few details about what they actually use.

What to ask instead

  • Where does lineage break?Which of our sources, transformations or tools will produce gaps?
  • Who approves the AI text?Is generated content marked, reviewed and owned before users see it?
  • Show us our fiveThe connectors for our five main systems, live, with what each extracts.
  • Can we call a reference?A customer of similar size, sector and use case, without the vendor present.

The AI line deserves a word of its own. Generated descriptions can save a lot of effort, and I would not dismiss them. But a description that sounds right and is wrong is worse than no description, because people trust it. The question is not whether the tool can generate text. It is whether your process makes a named person accountable for each description before it is published as fact.

A generated description that sounds right and is wrong is worse than an empty field, because people believe it.

Bernard Millet, DataEdgePro.net

Part VII · The Decision

Scoring, Total Cost and the Exit

When the proof of concept is done, score each product against the criteria you weighted in Part IV, using what happened during testing. Keep the scoring sheet short. Six to eight criteria, weighted by your primary use case, are enough. Longer sheets tend to reward the product with the most features rather than the one that fits.

Then look at total cost over three years, not the first year price. Licence is only one part. Add the implementation effort, the hosting and operations if you self host, the integration work for your less common sources, and above all the curation time of your stewards. The last item does not appear on any vendor quote, which is exactly why it is the one most often forgotten.

Finally, plan the exit before you enter. Catalogs tend to stay for years, and the curated content inside them, the descriptions, glossary links, owners and certifications, represents a large investment of your people's time. Make sure the contract and the API let you take that content with you. The DAMA-DMBOK treats Metadata Management as one of its eleven knowledge areas, a discipline in its own right rather than a feature of a tool.[5] Your metadata belongs to you. The catalog is where it lives for now.

Three Signs Your Catalog Will Become Shelfware

One of these deserves attention before signing. Two or more means the problem is not the tool you choose.

No steward
Nobody has time in their objectives to curate the first domain, and nobody has been asked.
No question
The project cannot name five real questions users will answer with the catalog in its first year.
No first domain
The plan is to connect everything at once and curate later.

Part VIII · The Market in 2026

Four Families of Catalogs in 2026

Once the method is clear, the natural next question is where to start looking. The market moves fast, through acquisitions, rebrands and new features, so rather than a ranking, here is a map. As of October 2026, catalogs fall into four broad families. Each fits a different situation, and the right family for you usually follows from your existing platform and your primary use case.

FamilyExamples as of October 2026, alphabeticalUsually fits when
Open sourceDataHub,[6] OpenMetadata,[4] Unity Catalog (open source edition)[7]You have engineering capacity, want full control of hosting and metadata, and want to avoid licence lock in. Managed versions are also offered by the companies behind some of these projects.
Standalone enterprise catalogsAlation, Alex Solutions, Atlan, Coalesce, Collibra, DataGalaxy, data.world (now part of ServiceNow), OvalEdge, Solidatus[8][9][10]Your estate spans many platforms, governance and stewardship workflows are central, and you want one catalog above all your systems.
Built into a data platformAmazon SageMaker Catalog,[11] Databricks Unity Catalog,[7] Google Dataplex Universal Catalog,[12] Microsoft Purview Unified Catalog,[13] SAP Datasphere catalog,[14] Snowflake Horizon Catalog[15]Most of your data already lives on one platform, and access control, lineage and discovery close to the data matter more than coverage of other systems.
Part of a data management suiteAb Initio, Ataccama, IBM watsonx.data intelligence,[16] Informatica (now part of Salesforce),[17] Precisely, Quest Software[10]You already use the vendor's integration, quality or master data tools and want the catalog tightly connected to them.

Three practical notes. First, if most of your data sits on one cloud data platform, test its built in catalog before buying anything else. It may cover your first use case at little extra cost, and it gives you a baseline for every other demo. Second, the families overlap. Standalone catalogs connect to the platforms, several platform catalogs now include business glossaries,[11][12][13] and some open source projects have commercial editions.[6] Third, ownership changes matter. Two well known names in this list have been part of recent acquisitions: ServiceNow announced the acquisition of data.world in May 2025, and Gartner now lists it as ServiceNow (data.world),[9][10] and Salesforce completed its acquisition of Informatica in November 2025.[17] When a vendor changes hands, ask how the roadmap, pricing and support for your use case will be protected.

How complete is this list? No list of a fast moving market can be exhaustive, but this one includes every vendor evaluated in Gartner's Magic Quadrant for Metadata Management Solutions, published on 19 November 2025,[10] plus the open source projects and platform catalogs, which are not among the vendors evaluated in that report. Gartner's separate Magic Quadrant for Data and Analytics Governance Platforms, published in January 2025, also includes Microsoft and BigID, a data security and privacy platform with discovery and classification features.[18] Analyst reports are a useful way to build a long list. They are no substitute for testing on your own data.

None of these families is better in general. Run the same questions, the same proof of concept and the same red flag checks across whichever two or three you shortlist, and let your own data decide.

Conclusion: Choose for the Second Test

Every serious catalog on the market can pass the first test, the demo. The choice is about the second test, a year later, when a business user searches for something that matters and either trusts what they find or does not. That outcome depends on the fit between the tool and your use case, on how easy curation is for your people, and on whether someone owns the content.

So settle the use case, the first domain and the curation model before you call a vendor. Ask for evidence on your own systems. Test with your own people. Weigh the criteria that matter for your use case, count the full cost and keep the exit open.

If I had to leave one sentence with a CDO about to buy a catalog, it would be this: pick the tool your stewards can live with, because they are the ones who will make it worth paying for.

Sources & references

  1. Federal Act on Data Protection (FADP, SR 235.1), in force since 1 September 2023, article 12, Records of processing activities. Required content of the record, and exceptions for legal entities with fewer than 250 employees whose processing poses a negligible risk. lawbrary.ch · official text: fedlex.admin.ch
  2. MME, Totally revised FINMA Circular 2023/1 "Operational Risks and Resilience, Banks". Entry into force on 1 January 2024, expanded supervisory practice on handling critical data. mme.ch
  3. MME, Cloud services at banks and securities firms, part 1, 3 December 2025. The institution remains responsible for the proper performance of functions outsourced or transferred to third parties. mme.ch
  4. OpenMetadata, project repository. Released under the Apache License, Version 2.0. github.com
  5. DAMA International, DAMA-DMBOK: Data Management Body of Knowledge, 2nd edition, Technics Publications, 2017. Metadata Management as one of eleven knowledge areas in the DAMA Wheel. dama.org · summary of the knowledge areas: datacrossroads.nl
  6. DataHub, project repository. Originally built at LinkedIn, open sourced in 2020, released under the Apache License 2.0, with a commercial DataHub Cloud offering. github.com
  7. Databricks, Open Sourcing Unity Catalog, 13 June 2024. Unity Catalog released under the Apache 2.0 licence and hosted by LF AI & Data. databricks.com
  8. Vendor websites: alation.com · atlan.com · collibra.com
  9. TechCrunch, ServiceNow acquires Data.World months after snatching up Moveworks, 7 May 2025. ServiceNow signed a definitive agreement to acquire data.world, described as a cloud native data catalog and data governance platform. techcrunch.com
  10. Gartner, Magic Quadrant for Metadata Management Solutions, 19 November 2025, by Melody Chien, Thornton Craig, Guido De Simoni and Roxane Edjlali. Fifteen vendors evaluated: Ab Initio Software, Alation, Alex Solutions, Atlan, Ataccama, Coalesce, Collibra, DataGalaxy, IBM, Informatica, OvalEdge, Precisely, Quest Software, ServiceNow (data.world) and Solidatus. gartner.com
  11. Amazon Web Services, Amazon SageMaker Catalog. Built on Amazon DataZone, includes a business glossary. aws.amazon.com
  12. Google Cloud, Standardize your business terms with Dataplex business glossary, now GA, 11 June 2025. cloud.google.com
  13. Microsoft Learn, Microsoft Purview Unified Catalog. Glossary terms, data quality and data products. learn.microsoft.com
  14. SAP Learning, Using the Catalog and Marketplace Features in SAP Datasphere. Central catalog to discover, classify, profile and understand enterprise data, with business glossary, KPIs and impact and lineage analysis. learning.sap.com
  15. Snowflake documentation, Snowflake Horizon Catalog. Discovery, governance and lineage. docs.snowflake.com
  16. IBM, watsonx.data intelligence. Enterprise data catalog with business definitions, ownership, quality and lineage. ibm.com
  17. Salesforce, Salesforce Completes Acquisition of Informatica, 18 November 2025. investor.salesforce.com
  18. Gartner, Magic Quadrant for Data and Analytics Governance Platforms, January 2025. Fifteen vendors evaluated, including BigID and Microsoft. Vendor list: aidataanalytics.network · publication date: ataccama.com