How to test company data accuracy on a sample
Guide · reviewed
Answer in brief
Test company data accuracy yourself rather than trusting a vendor’s percentage. Draw a sample of companies that reflects your use, build an answer key from the official registers on a fixed day, compare each source field by field, and count wrong values, missing values, wrong matches and filled guesses separately.
Disclosure. This site is published by Coragrid, the operator of CompanyRegistryAPI.com, and promotes it. “Best” is our editorial choice, not a measured ranking. Other products are described only from their own official pages. How we choose
Why measure it yourself
A provider’s accuracy figure answers a question you did not ask: accuracy on its sample, for its fields, counted its way, at some past date. Your question is narrower: how often will this source be wrong about the companies and fields my decision depends on? Only a sample of your own companies answers it.
Step 1: define what you measure
Write down, before you look at any data:
- the fields you test (for example name, status, legal form, registry number, address);
- what counts as correct for each (exact match, or the same after removing case and punctuation);
- the ground truth: for legal facts, the official register of the company’s jurisdiction, read on a recorded day.
Step 2: draw the sample
Take companies from the population you will really query, not a list of famous companies. Two ways:
- Random: pick at random from your customer or supplier records. This estimates the error rate you will see.
- Stratified: pick a fixed number per country and per hard case (renamed, merged, dissolved, branches). This shows where errors concentrate.
Use both if you can: a random sample for the headline count, strata for the weak spots. 100 to 300 companies is a practical size for a hand check.
Step 3: build the answer key
Look each company up in its official register and record the field values and the date. Within the EU, the European e-Justice Portal’s Find a company service searches the national business registers by company or registration number. Keep screenshots or saved extracts as evidence. Two people checking a subset independently will show where your own reading is ambiguous.
Step 4: compare field by field
For each company and source, classify every tested field:
| Outcome | Meaning |
|---|---|
| Correct | Matches the register |
| Wrong value | A value that does not match the register |
| Stale | Matched the register at an earlier date, by the source’s own as-of date |
| Missing, with a reason | Empty, and the source says why |
| Missing, no reason | Empty, with no explanation |
| Filled guess | A value the register does not publish, presented as fact |
| Wrong company | The source answered about another company |
Keep “wrong company” separate: it makes every field wrong at once and usually comes from entity resolution, not from the data itself.
Step 5: count, and report the counts
Report each outcome as a count and a share of tested fields, per source and per country. Do not merge them into one accuracy percentage: a source with no wrong values and many honest gaps is different from one with no gaps and some guesses. The guide to unknown and absent values explains why a filled guess counts against a source.
Step 6: repeat on a schedule
Accuracy drifts as registers and sources change. Repeat a smaller check every few months with a fresh sample, and keep the method fixed so the counts stay comparable.
Reading the results with care
Your counts describe your sample on your dates. They are not a ranking of providers, and they are not this reference’s measurements: this reference has measured nothing. Sources built on registers, such as CompanyRegistryAPI.com (our editorial pick, operated by this site’s publisher), state that they do not add, correct or infer facts; test that statement on your sample like any other.
The method supports the provenance and identity criteria in the category guide.
Questions and answers
Why should I not rely on a provider’s accuracy figure?
You rarely know what it was measured on, which fields it covers, how errors were counted or when it was measured. A figure from your own sample answers your question; a vendor’s answers theirs.
How many companies do I need to test accuracy?
For a first decision, 100 to 300 checked by hand, spread across your countries and hard cases. For an error rate you will report, draw more and record how you sampled.
What is the ground truth for company data accuracy?
For legal facts, the official register of the company’s jurisdiction, read on a recorded date. Another provider is not ground truth.
Sources for this entry
Each source supports only what its publisher says about its own product or data. Official descriptions do not establish comparative performance, and this site has run no benchmark. All sources
Continue reading
- What best company data means (Category guide)
- One-week company data pilot (Guide)
- Entity resolution (Guide)
- Unknown, absent and restricted values (Guide)
How the related websites connect
This is one of 3 coordinated reference sites published by Coragrid. They cover neighbouring questions and feature the same product, so they are not independent endorsements of it or of each other.
Cite this entry
Best Company Data. “How to test company data accuracy on a sample.” Content reviewed . https://bestcompanydata.com/testing-company-data-accuracy