What is Company Data With Sources?

Provenance is not the same as a source list

Almost every data provider names sources. Usually that means the origin of the whole dataset: “from the commercial register, company websites and public directories”. That is a statement about the pot, not about the single record. A figure only becomes testable when the evidence is attached to the statement itself.

The test for this is trivial and still splits the market: Take any row from the export and ask where exactly this one statement comes from. If you get back an address where the sentence still sits today, that is evidence. If you get back a list of source categories, that is provenance.

The three parts every piece of evidence needs

• The quote. The sentence carrying the statement, word for word as it appears at the source. A summary is not enough: you cannot hold it against the page.
• The location. The concrete address, not the provider's homepage or a portal. A link to the overview page of a news site is no evidence for a story published two weeks ago.
• The date. When the statement was published and when it was read. Without a date, “the company is hiring a production manager” is worthless: the sentence could be from yesterday or from 2019.

Which providers carry sources per record

The most reliable group are register platforms. Anyone taking data straight from official registers can trace every field back to the registering authority, because exactly one document sits behind it: legal form, registered office, officers, annual accounts. That holds for credit agencies and register portals as much as for the registers themselves, which you can query for free.

The limit of this group is content: a register knows who the company is, not what it does. Manufacturing processes, machinery, certifications, ongoing investments, open positions: none of that sits in a register. Those statements live on the company's own website, in its press releases and in tenders, and that is exactly where most datasets harvest them without recording where they came from.

Why this hurts most with contact data

For people, a missing source is not just inconvenient, it is legally relevant. An email address derived from a pattern (“firstname.lastname@”) looks exactly like one that appears in a legal notice. The difference decides whether you are processing a documented figure or a guessed one. A dataset that makes both look the same takes that decision away from you without asking.

The same trap applies to currency: a derived address carries no date. Whether the person still works there cannot be read from the record as long as nobody asked the question and wrote down the answer.

How to test a provider in five minutes

• Ask for ten rows and pick three statements that are not register facts: a process, a certification, an investment.
• Ask for the source of each one. Open the address and look for the sentence on the page.
• Check the date. Is there one? Does it match what the page says?
• Take one email address and ask whether it was found or derived. “Our system verifies all addresses” is not an answer to that question.
• Ask for something the provider does NOT have. A dataset that returns something for every question does not distinguish between “looked, found nothing” and “never looked”.

The uncomfortable consequence

A documented dataset is smaller than an undocumented one. If every statement is bound to a source, fewer rows survive, because the rows nobody has evidence for drop out. That is not a weakness of the method but its result: you see what is actually known, instead of a filled-in table whose gaps only show up on the call.