An Asian company database for CRM: removing duplicates while retaining sources
Combine company lists from Japan, Korea and China without losing names, sources or relationships between entities. Practical acceptance rules for a database used by sales.
Article language: English

In this article
Two hundred rows in a spreadsheet do not necessarily mean two hundred companies. The list may include a brand, its owner, a company's former name and several branches described as independent organisations. After import into a CRM, a system for managing customer relationships, those differences begin to affect contact assignments, market assessment and sales planning.
Preparing a database means identifying what each row represents and retaining evidence for that decision. It is not a full assessment of a partner's trustworthiness or permission to contact them. This guide concerns the quality of organisation data: identities, names, relationships, sources and merge rules. These help the team understand whom the information refers to.
First define the unit represented by a record
Decide whether you are collecting legal entities, brands, manufacturing sites, shops or corporate groups. All may be useful, but they should not appear to be the same kind of object. If the goal is a conversation about group purchasing, a local shop may be a source of information without necessarily being the decision-making unit. This should be visible in the database before a salesperson is assigned a task. For a database of Korean facilities, also consider the role of industrial complexes in the audience map (in Polish). A head-office address, a production site and a partner's area of operation may describe different aspects of the same record.
Put the definition in the project instructions. For example, the main record describes an organisation, with its brands and locations as related elements. When researching laboratories in China (in Polish), also connect the right legal entity, the location and the scope of the particular test; one group record cannot explain all three. If the CRM lacks suitable relationships, add controlled fields for record type and parent-organisation identifier. Do not put a group name in the city field simply because it is empty. That time saving quickly becomes a repair cost.
Agree the basic model before collecting data. In your own research brief, specify the unit, fields and delivery format. Our guide to preparing an Asia market research brief (in Polish) explains how to define the scope. If the provider is to deliver a list of brands while sales expects decision-making entities, even carefully collected material will not meet the process's needs. That scope difference cannot be fixed merely by removing repeats.
Keep the original name and spelling variants
Store the name in its original script and, separately, the name used in international communication if confirmed. Latin-script variants can differ between directories. Do not automatically choose the first search result as the only correct name. Record the source and whether the entry is an official spelling, a transliteration or an analyst's working label.
Normalisation helps compare data but should not destroy the source content. You can prepare a helper field without surplus spaces or differences in letter case. Do not use it to replace the displayed name. Abbreviations, characters and name components can help distinguish entities. A rule that removes them from every record can increase apparent matches rather than reduce them.
Keep addresses and domains in source and working forms too. A shared domain does not prove there is one company: it may serve a whole group. A similar address may be an office building with many organisations. Merging needs a combination of evidence, not one impressive name-similarity result. Doubtful cases should go to manual review.
Choose identifiers for the country and entity type
If an appropriate public organisation identifier exists, record it with the system name and source. In Japan, the National Tax Agency publishes basic information including the name, address and Corporate Number. It also explains that the number is assigned to an organisation, not its branches. This is useful when separating an entity from a location. NTA's description of the Corporate Number.
Do not apply that model mechanically to Korea or China. Check which number occurs in each source and what it means. Do not mix a registration identifier, stock ticker and internal directory number in one field. Mark an identifier not found as unconfirmed within the checked scope. Do not automatically infer that the organisation does not exist.
For a record requiring further assessment, use the procedure for verifying a partner in Korea, Japan or China (in Polish). Orderly data helps start that work but does not replace checking the claimed activity. A record can identify the organisation correctly and still be unsuitable for the intended collaboration or have an outdated contact.
Create merge rules and an exception list
Divide matches into three groups: confirmed, probable and unresolved. The first may follow from the same appropriate identifier and consistent data. The second needs further checking, for example when names are similar and an address is shared. The third stays separate until clarified. The instructions should say who may approve a merge and which sources can support it.
Before merging, choose a primary record and decide how fields will be retained. The newest file does not always contain the freshest information. Material collected today may quote an old directory. Keep the source content date if known and the date of your own check. When data conflicts, do not hide it in an overwritten cell. Record a task to resolve it.
A merge decision record should identify:
- The input records and their identifiers in the original files.
- The reason for treating them as the same object and supporting sources.
- The destination record and how contacts and notes will be transferred.
- Data rejected, retained as historical or still disputed.
- The person approving the decision and its date.
- The ability to restore the state before the operation.

A hypothetical scenario: three lists of Japanese integrators
In this working example we combine a trade-fair directory, an earlier sales list and new research results. We use no real customer data. One company appears under a Japanese name, an English name and a brand name. Its two offices also appear separately. Simply removing similar names might delete useful location information.
The analyst identifies the main entity, retains name variants and links its branches to it. The brand gets a separate record type linked to the organisation. Two ambiguous entries remain separate with a question to resolve. The salesperson can see which data is ready for use and which should not yet be interpreted as a new company in the market.
After cleaning, the list is shorter, but that does not mean the market is shrinking. The unit of counting has changed. Explain that distinction to the reader of your own market report. The guide to reading a Korean market report (in Polish) discusses how to interpret such material. Otherwise better data quality may be mistaken for a sudden fall in potential customers or a provider's error.
Accept the database on a sample before importing everything
Prepare difficult cases: the same name for different companies, one entity with several brands, a changed domain, a branch and a record with no identifier. Check whether the instructions lead two reviewers to similar decisions. Differences will reveal where clarification is needed. This test is more valuable than assessing only easy, complete rows.
Run a small trial import and a control export. Check original characters, relationships, field lengths, source addresses and the handling of empty values. Make sure missing information has not become false text or a default category. Compare record counts by type, not just the total number of rows. Record how to undo the trial before expanding the operation.
After import, test specific user tasks: finding a company by its original name, checking its brand, identifying the information source and recognising the contact owner. If those actions require opening three private spreadsheets, the database is not yet ready for independent use. Acceptance should measure usefulness, not only the absence of a technical error during file loading.
Set updates and boundaries for data use
For each field needing review, identify an owner or event that triggers a check. It might be a bounced message, a change to an organisation's site or preparation for an important conversation. Not every record needs refreshing equally often. In East Asian market monitoring (in Polish), account for signals that call key information's currency into question.
A competitor-analysis database needs different fields from one for handling enquiries. Products and channels may matter for the first, as explained in competitor analysis in Korea (in Polish). For the second, the contact's role and current case stage matter. Do not add personal data without a specific operational purpose. The quality of an organisation record does not settle the conditions for using every related contact.
Once identities are organised, move on to selecting partners for a conversation (in Polish). This is the next stage: a clean record does not yet establish a company's suitability. Separate the data-check result from the reasons an organisation was shortlisted.
Assess quality without false precision
Before acceptance, agree separate measures for completeness, correctness and duplicate resolution. A record may have every required field and still describe the wrong company. It may also have a correctly established identity but no public contact for the required role. One combined “database quality” score hides these differences and makes it harder to tell the provider what to improve.
Choose a sample covering different sources and record types. Record the selection method so you do not assess only entries prepared by one person or taken from the easiest directory. For each, check identity, object type, the source of a key field and the justification for a merge. Treat assigning information to the wrong entity separately from a small correction to a name's spelling. Their effects on the team's work differ.
If you find a problem, examine its rule. An error in one row may result from a manual oversight, but it may also reveal flawed normalisation across the list. In the latter case, review all records processed by the same method. Correcting the identified cell alone does not resolve the cause. The provider should describe the scope of the check and confirm which entries were reassessed.
Finally, retain a reconciliation: the number of input records, confirmed merges, objects by type and unresolved cases. Do not add these categories without checking whether they are mutually exclusive. Include a list of gaps deliberately left open. The recipient should be able to distinguish limitations of the available source from an unfinished part of the commission. That gives further work a clear scope instead of an implicit expectation that every empty cell will someday be filled.
A properly accepted database makes it possible to say what was counted, how it is known, which records were merged and what remains uncertain. Those four answers matter more than an impressive row count. If you want a database for a specific process, present the task to Aeneia Oriens.








