What Claude Cites When You Ask It About Suppliers

"Claude doesn't really look things up. It answers from memory, so there's nothing to learn from its sources." That objection is reasonable, and we made a version of it ourselves in an earlier piece on brand monitoring in Claude, where we described an assistant that leans heavily on what it absorbed in training. On 20 October 2026 we checked it against our own data: what Claude cites when the question is about suppliers, how often, and from which kinds of site. Part of the objection fell apart. Part of it survived.

What we counted, and on which answers

PSentry keeps every answer it collects. Between 13 December 2025 and 20 October 2026 that came to 17,719 answers from ChatGPT, Gemini, Claude and Perplexity, all gathered through the providers' APIs, with 166,524 source URLs attached. Claude accounts for 3,904 of them: 1,606 answers to prompts in Italian, 1,156 in French, 1,076 in English and 66 in Spanish.

We counted three things. How many answers carried at least one source. How many sources each answer carried. And what kind of site those sources were: a country domain or a .com, a B2B directory, Wikipedia, a social platform.

The limit belongs up front, because it shapes everything after it. The prompts are the ones PSentry customers monitor, mostly about B2B services and industrial products. They are supplier questions, not a sample of everything people ask Claude. There are no German prompts in the set at all, and 66 Spanish answers are too few to read anything into. What follows describes English, Italian and French.

Claude cited sources in 97 answers out of 100

The first figure undid the objection. Claude returned at least one source in 97 answers out of 100. Perplexity, the assistant people associate with visible sources, did so in 94. Gemini did in 56 out of 100, ChatGPT in 44.

Claude also carried the most sources per answer: 12.4 on average, against 11.6 for Perplexity, 10.3 for Gemini and 3.9 for ChatGPT.

Anthropic's documentation for the web search tool explains why a supplier question behaves this way. When the tool is available, Claude decides for itself when to search, and the documented triggers include information about specific organisations, people or products that might have changed. "Who makes stainless steel valves for the food industry in Northern Italy" is exactly that kind of question. The same page states that citations are always enabled for web search. When Claude searches, the sources come with the answer by design.

This is not a finding about the Claude app. Our answers come through the API, and PSentry queries Claude Haiku 4.5, the compact model Anthropic released on 15 October 2025. The app a buyer opens may run a different model, with its own instructions. What holds is narrower: put a supplier question to Claude through the API with search available, and a source list is close to certain.

More sources, drawn from fewer sites

The second figure is where the objection survived, in a different form. Claude cites a great deal, but it keeps returning to the same places.

In English, Claude's answers carried 11,381 sources from 1,664 distinct domains. Perplexity's carried 15,652 sources from 4,890 domains. Claude cited each domain about seven times on average. Perplexity cited each about three times. Italian and French had the same shape: roughly eight citations per domain for Claude in Italian against fewer than four for Perplexity, roughly seven against three in French.

For a supplier this cuts both ways. A narrow circle is harder to get into. A site that is inside it turns up across many answers, not one.

B2B directories weigh more, in Italian above all

By directories we mean the B2B and review catalogues: Europages, Kompass, Sortlist, Clutch, G2, Trustpilot, Capterra, GoodFirms and DesignRush. In Italian, about 6 of every 100 Claude sources were one of these. Perplexity was at about 4, Gemini about 2, ChatGPT under 2. In French Claude came first again, at just under 4 in 100.

English broke the pattern. There Perplexity cited directories more often than Claude, about 3 sources in 100 against 2. So "Claude prefers directories" is too strong a summary. In our sample it leaned on them most when the question was in Italian or French. We do not know why English differs, and the counts cannot tell us. Why these catalogues get cited at all is the subject of our piece on G2 and Trustpilot.

Social platforms went the other way. YouTube, LinkedIn, Instagram, Facebook and Reddit together made up less than one Claude source in 100, in all three languages. Perplexity drew between 3 and 7 in 100 from them. Wikipedia stayed marginal for Claude everywhere, peaking at just over one source in 100 in French.

The language of the question chooses the country

In Italian, 55 of every 100 Claude sources were .it domains and 35 were .com. In French, 36 were .fr and 51 were .com. In English, 80 of every 100 were .com, the highest share of any of the four assistants, and fewer than 4 were .uk.

For an exporter this matters more than any average above. The same buyer question in two languages draws on two different webs. A company that is thoroughly documented on .com sites is not thereby present in Claude's Italian answer, which reads mostly Italian sites.

What this test did not show

• Anything about German. No German prompts, no German answers. Any claim about Claude's sources in German would be invented.
• Anything reliable about Spanish. 66 answers is an anecdote, not a pattern.
• How the Claude app behaves. Different models, different instructions, a different surface.
• Why a site is in the circle. Domain counts say where Claude looked, not what it valued there.
• A trend. These are totals over ten months, and the assistants changed during them.

Running the same check on your own category

The test is small enough to repeat by hand. Write five questions a buyer would ask before knowing your name, in each language you sell in. Put them to Claude with web search on and copy every source it lists. Then sort the list into five columns: your own site, competitors, directories, trade press, everything else. If the directory column is long and your company is missing from the directory pages Claude cites, that is the first gap worth checking.

PSentry does the same across ChatGPT, Claude, Gemini and Perplexity, twice a month, per language and market, and keeps the sources each assistant used. It shows where Claude looked. It cannot make Claude look at you.