How to check an AI-generated distributor list before you trust it
Five checks for any AI-made distributor list — notice link, distributor or manufacturer, registry, product fit, lawful channel — and why a chat answer differs.

Tomasz Kowal · Data and public procurement
Updated 29 October 2026 · 5 min read

Twenty confident company names are the easiest thing for a language model to produce. Whether they exist, still trade, resell your kind of product and may lawfully be e-mailed is another question. These are the five checks to apply to any list, including ours.
Putting it simply: Before you trust an AI-generated distributor list, check that each entry links to a real document, that the company resells rather than manufactures, that it is still registered, that its range fits your product, and that the contact channel is lawful in that country. A list without those five things is a draft, not a result.
Why lists need checking
Generative models produce what NIST calls confabulation: "confidently stated but erroneous or false content" — hallucination, colloquially. A 2024 study in Nature put it plainly: large language models "fluently make claims that are both wrong and arbitrary". A company name is a claim like any other.
The remedy is not a better prompt but evidence: a document that exists independently of the model.
Check 1: Is there a link to a notice?
In Cresques the source behind every entry is a contract award notice on TED, the Supplement to the Official Journal of the EU, which publishes more than 3,000 procurement notices each working day. Public buyers above the EU thresholds must send such a notice within 30 days of concluding a contract; it names the winner, its address, whether it is an SME, the value and the CPV codes (Directive 2014/24/EU, Article 50 and Annex V Part D).
Open the notice. Does the name match? Is the buyer real? Is the date recent enough to matter? If there is no link, or it does not lead to a document naming the company, treat the entry as unverified.
Check 2: Distributor or manufacturer?
A company that won a tender in your category is either a maker (your competitor) or a reseller (your potential partner). The notice rarely says which.
In a Cresques company profile, a language model researches that label from the notices and the company's public website: independent distributor, manufacturer, manufacturer's subsidiary, or both. Each label comes with verbatim quotes checked on the source pages; when no quote can be confirmed, the role stays "unknown". In our internal tests on a sample of notices (see our methodology), roughly 7% of automated labels were still wrong after corrections — so the label is a sorting aid, not a verdict. One minute on the company's website — its own factory and brand, or a catalogue of other people's brands? — is the check.
Check 3: Is the company still trading?
An award notice is a snapshot. The company may have renamed, merged, moved or closed since.
Use the official register, not a directory. The European e-Justice Portal's "Find a company" service searches the business registers of all EU countries plus Iceland, Liechtenstein and Norway through the Business Registers Interconnection System. The portal shows what each national register provides free of charge; paid documents such as annual accounts are ordered from the national register itself. Confirm name, status and address before you write.
Check 4: Does the product fit?
CPV codes are deliberately broad. The Common Procurement Vocabulary is one EU classification (eight digits plus a check digit, mandatory in notices), and a code such as office furniture covers desks, chairs, storage and more. A distributor that won a contract for seating is not necessarily interested in your acoustic panels.
Read what the notice says was supplied, then look at the company's range. Would your product sit naturally next to what they already sell? If the honest answer is "not really", move the company down the list.
Check 5: Is the contact channel lawful?
An e-mail address next to every company invites you to press send. Whether you may depends on the country.
In France, CNIL allows e-mail prospecting of a professional without prior consent, provided the message relates to their professional function, they were told they could object when the address was collected, and every message lets them opt out. In Germany, § 7(2) No. 2 of the Unfair Competition Act (UWG) treats advertising by electronic mail without the recipient's prior express consent as an unreasonable nuisance — so a cold e-mail is the wrong first channel there.
So: what is the lawful first contact in this country, and does the list tell you? Cresques shows the route per market; a list that does not leaves the legal risk with you.
A chat assistant versus an evidence-backed list
The difference is where the names come from, not whether AI is involved.
Ask a general chat assistant for "furniture distributors in France" and it generates names from patterns in its training data: no document, no date, no buyer, no value. Some names will be right, some mixed up with other companies, some non-existent — all in the same confident tone. You cannot tell which is which without doing all five checks from scratch.
An evidence-backed list starts the other way round: with a notice that exists, naming a company that won a specific contract on a specific date. The model's job shrinks to labelling and sorting what the document already says. It can still be wrong about the label — see check 2 — but it cannot invent the company, and every claim arrives with its source. The checks become quick confirmations rather than an investigation.
What this means for a manufacturer
- Treat any distributor list, from any tool, as unverified until each entry links to a real document.
- Budget a few minutes per company for the five checks — an afternoon, versus months of e-mails to the wrong people.
- Keep the registry check even when the evidence is strong — notices age.
- Decide the contact channel per country before you write a single message.
- Prefer tools that show their source and their uncertainty over tools that only sound sure.
