Your supplier list holds Hartog Industrial Supplies, Hartog Industrial Supplies B.V. and Hartog Industrial Suplies. All three are in Rotterdam, and two carry the same VAT number. A unique rule on Name let all three in, because no two names are the same text.
So you search for data matching software, and every demo shows the same screen: pairs of records with a score. The score is the easy part. The questions that decide whether a tool is any use come after it. Which records did it skip? Who presses Merge? What happens to the purchase orders that point at SUP-1002? Can you take the merge back?
We built Lookalikes, the matching in Primentra, against the questions below, so we also say where Primentra stops. Ask them in any demo, ours included.
Two kinds of data matching software
Data matching software finds records that describe the same real thing when their values differ. It comes in two kinds, and buying the wrong kind is the first expensive mistake.
Many standalone tools work on a copy. You export the list, the tool matches it, and you load a clean file back. That is fine for a mailing list. For master data it leaves you the hard part, because your orders and child records still point at the codes you just removed.
The questions to ask in the demo
1. Why did these two records score 91?
You cannot check a score that comes without a reason, so ask for the reason per column. In Primentra a pair reads like this: Name 99% similar, City exact, VAT number empty on one side, not counted. With that line in front of them, a steward decides in seconds. Given only the 91, they open both records side by side and compare them field by field.
2. What does an empty column do to the score?
Ask this with a real pair where one side is empty. Some tools count an empty value as a mismatch, and then they miss each duplicate where somebody skipped a field. Others count it as a match, and then two records with no VAT number look alike on VAT. In Lookalikes an empty column does not count at all. It does not raise the score and it does not lower it.
3. Which records did it never compare?
No tool compares every record with every other record. At 50,000 suppliers that is about 1.25 billion pairs. So a tool first groups records by a cheap key, such as the first letters of the name, and compares only inside a group. That step is blocking, and our post on how matching works explains it.
A record that lands in no group is never compared, and a clean result screen does not tell you about it. Ask to see that list. Primentra has a Not checked tab for it. It shows records with no value to compare, and records in a group larger than the block cap, which is 1,000 by default.
4. Does it remove B.V. from a name, and keep BV in a postcode?
Legal forms make one company look like two: B.V., GmbH, Ltd, S.p.A. A tool must strip them before it compares names, and leave them alone everywhere else. In the Dutch postcode 3511 BV, BV is part of the address. Lookalikes removes legal forms from the Name and from columns compared for typos, and nowhere else. You can add the forms of your own countries to the list.
5. Who presses Merge?
Auto-merge sounds like a time saver until it merges two subsidiaries with different VAT numbers. Ask what must be true before the tool merges on its own.
In Primentra, auto-merge is off until an administrator switches it on. Even then a pair needs a score in the auto band and at least two columns that count, carrying at least 70% of the weight. A name and a city alone never qualify, which is why the Hartog pair above goes to a person at 99. On an entity that needs approval, auto-merge only proposes the merge, and an approver decides.
6. What happens to everything that pointed at the record that went away?
Vendors skip this question in most demos. A supplier code lives on in purchase orders and in the ERP vendor table, and a merge has to settle each place that holds it.
In Primentra each domain link and child record moves to the kept record, in every model. The old code keeps working. The picker answers SUP-1002 with the kept record and says that SUP-1002 was merged into SUP-1001. A link by the old code in an import also goes to the kept record. A staging row that uses the old code as its own Code fails with error 524288, and the REST API answers 409 CODE_MERGED_AWAY. So a source system that still sends the old code cannot overwrite the merge.
7. Can you undo a merge next week?
Some merges are wrong, and you find out later. Ask whether unmerge gives you the records back, or only a line in a log. Unmerge in Primentra brings back the merged records with the same Id, Code and values. A link comes back when it still points at the kept record. A change a person made after the merge stays.
8. Can you test a rule on your data without changing it?
A matching rule is a set of weights and thresholds, and the first draft is always wrong. Ask whether you can run it, read the result and run it again without writing anything. Test these rules in Primentra does that. It shows the pairs per band, the groups and the records not checked, with a reason for each score, and it saves nothing.
9. Where does your data go?
Matching tools now offer an AI opinion on a pair. Ask which fields leave your network, and to whom. In Primentra, AI assist is off by default and gives advice only. It sends the compared columns of one group to the provider and key you set up, and never a column the user may not see. That provider can be a model server on your own network.
SQL Server 2025 can score a pair. You build the rest.
If your master data lives in SQL Server, you may ask whether you need software at all. SQL Server 2025 added four fuzzy functions: EDIT_DISTANCE, EDIT_DISTANCE_SIMILARITY, JARO_WINKLER_DISTANCE and JARO_WINKLER_SIMILARITY. They are in preview, so you switch them on per database first.
ALTER DATABASE SCOPED CONFIGURATION
SET PREVIEW_FEATURES = ON;
-- every pair of names, 0 to 100
SELECT TOP (100) a.Code, b.Code,
JARO_WINKLER_SIMILARITY(
a.Name, b.Name) AS Score
FROM dbo.Supplier AS a
JOIN dbo.Supplier AS b
ON a.Code < b.Code
ORDER BY Score DESC;That query scores one column and stops. Nothing groups the records first, so on 50,000 rows it scores about 1.25 billion pairs. The functions also refuse an nvarchar(max) column. For a one-off look at a small list it is a good start. Everything after the score is still yours to build: the legal forms, the review queue, the merge and the moved references.
Microsoft removed its own matching tool in the same release. DQS went out with MDS, and few teams noticed.
Where Primentra fits, and where it does not
Lookalikes matches records inside one Primentra entity. It does not match across two entities, and it does not link records from several source systems into one golden record. If your duplicates come from a CRM and an ERP that each keep their own customer list, you need the first kind of tool, and the enterprise hubs do that job.
If your problem is one supplier, customer or product list that your team maintains, often a list that used to live in MDS, matching it where it lives saves the export and the reload. The Lookalikes page in the docs shows every setting and who may merge.
Bring five records of your own to every demo
A vendor demo uses data that matches well. Bring your own instead:
- Two records of one company, typed differently, one of them with the legal form.
- A third record of that company, with a typo and an empty column.
- Two records that look alike and are different companies, such as two subsidiaries with different VAT numbers.
Ask the vendor to score them and explain each score. Then ask them to merge the first three, show what happened to an order that pointed at one of them, and undo the merge. Five records fit on one screen, so you can watch every step happen.
You can run the same test on Primentra. The trial runs for 60 days on your own SQL Server.
Common questions
What is data matching software?
Software that finds records describing the same real thing when their values differ, such as Acme Ltd and ACME Limited. It cleans a copy of the values, compares records in groups, gives each pair a score, and sends likely pairs to a person or merges them by rule.
Can Primentra match customers in our CRM against our ERP?
No. Lookalikes compares records inside one Primentra entity. To use it on a list from another system, load that list into an entity first, merge the duplicates there, and export the clean list.
Does Primentra merge duplicate records on its own?
Only when an administrator switches auto-merge on for that entity. On an entity that needs approval, auto-merge sends each merge to an approver as a request. Every merge can be undone with Unmerge.
Do I need SQL Server 2025 for Lookalikes?
No. Primentra computes the scores in its own service. It does not use the fuzzy functions that SQL Server 2025 added in preview.