Using ChatGPT, Claude or Gemini to find distributors: what they do well, where they invent, and how to check

"List twenty distributors of workwear in the Netherlands" is a tempting prompt. The answer arrives in seconds, formatted, confident and possibly partly fictional. That is not a flaw in one product; it is how general language models work, and all three major vendors say so in their own documentation.

Putting it simply: ChatGPT, Claude and Gemini are strong at synthesis, drafting and translation, and they can explain a regulation or structure a market plan in minutes. They are not registries. Asked for company names without a source document, they generate plausible text, and some of it may be wrong in ways you cannot see. Use them on evidence you already hold — a notice, a website, a register extract — and make them cite where every name came from. Then check each name yourself.

Disclosure: Cresques is our own product. We compare on criteria you can check yourself and link to each source.

What general AI assistants do well

Used for the right tasks, these tools save a manufacturer days:

• Synthesis. Paste a 30-page tender award and ask what was bought, by whom, for how much and under which CPV codes. Summaries of documents you supply are the strongest use.
• Drafting. A first version of an introduction letter, a distributor agreement outline, a product data sheet in a local format.
• Translation and tone. A French or German business e-mail that reads like one; a check of formality and idiom before a native colleague reviews it.
• Explanation. What a framework agreement is; how a buying group differs from a head office; what an Incoterm implies for your margin. Verify against the primary text, but the explanations are usually a good start.
• Structure. Turning scattered notes into a market-entry plan with steps, owners and dates.

None of these requires the model to know a fact it cannot show you.

Where they invent, and why

The US National Institute of Standards and Technology lists confabulation as a core risk of generative AI, defined as "the production of confidently stated but erroneous or false content (known colloquially as 'hallucinations' or 'fabrications') by which users may be misled or deceived". A 2024 study in Nature by Farquhar and colleagues describes how large language models "fluently make claims that are both wrong and arbitrary".

A company name is exactly the kind of claim that goes wrong. A model has seen many names of Dutch workwear suppliers in its training data and many patterns of how such names are formed. Asked to list twenty, it produces twenty strings that fit the pattern. Some exist. Some are real companies in the wrong sector. Some are plausible blends of two real names. The output does not distinguish between them, and neither can you without checking each one.

Two further points matter for distributor search. First, the model has no live view of a business register, a procurement portal or a company's current range unless a web-search feature is on and the model actually used it for that name. Second, the model does not know whether a company resells or manufactures, which is the single most important distinction for a manufacturer: one is a competitor, the other a prospect.

We have not run a test of these tools for this article and make no claim about how often any of them is wrong. The point is structural, not statistical.

What the vendors say, as of October 2026

The pattern across all three is the same: the vendor tells you to verify, and offers web search as a help. Web search makes a model better at finding pages; it does not make a page proof that a company resells your product, and it does not tell you whether you may lawfully e-mail anyone there.

A workflow that uses AI with evidence

The reliable order is evidence first, model second.

1. Start from a document that names companies. The strongest public source inside the EU is the contract award notice: when a public buyer concludes a contract above the EU thresholds it must publish one naming the winner, its address, the value and the CPV codes (Directive 2014/24/EU, Article 50). TED publishes more than 3,000 notices every working day, free.
2. Give the model the document, not the question. "Here are twelve notices; extract winner, buyer, CPV, value and date into a table." The model is now summarising, which it does well.
3. Ask it to label, and to say why. "For each winner, from the notice text and this website excerpt, is it more likely a manufacturer or a reseller? Quote the phrase you relied on." A quoted phrase is checkable; a bare label is not.
4. Ask for the source of every name. If the model adds a company it was not given, ask where it came from. If the answer is not a document you can open, drop the name.
5. Confirm in the register. The European e-Justice Portal's "Find a company" service searches the business registers of every EU country plus Iceland, Liechtenstein and Norway. Basic details are free.
6. Decide the lawful channel before drafting. Only then ask the model to draft the first message, in the local language, in the channel that country permits.

Prompts that help and prompts that hurt

How Cresques uses models differently

Cresques starts from the notice rather than from the model. You create a market as a country plus CPV categories; the companies shown are those named as winners in award notices, each tied to a notice you can open. In a company profile, a model researches the company's role — independent distributor, manufacturer, manufacturer's subsidiary, or both — and backs it with verbatim quotes checked on the source pages; if no quote can be confirmed, the role stays "unknown". That label can still be wrong, so you check it. What the model cannot do in this arrangement is invent a company, because every name enters through a document. Our materials reader works the same way in reverse: it reads your brochure or website and proposes CPV categories for you to confirm. Cresques also drafts a first message in the company's language for the lawful channel, which you send yourself, and e-mail alerts on new awards come with paid plans.

What this means for a manufacturer

• Use general AI assistants for synthesis, drafting and translation — tasks where you supply the facts.
• Never take a company name from a model without a document you can open behind it.
• Treat "distributor or manufacturer" as a label to check on the company's own site.
• Confirm every company in the business register; notices and models both age.
• Decide the lawful first-contact channel for the country before any message is drafted.