Country Domain or .com: Which Sites AI Assistants Cite

On 20 October 2026 we exported every source URL PSentry had collected since 13 December 2025: 17,719 answers from ChatGPT, Gemini, Claude and Perplexity, carrying 166,524 links between them. We wanted a plain answer to one question. When an assistant goes to the web to answer a buying question, does it cite a country domain or a .com, and does that change with the language of the prompt? It does, and by more than we expected. In English, three sources in four were .com. In Italian, about half were .it.

This is the log of that count: what we measured, what came out, and the parts of the sample too thin to say anything about.

What we counted

Every answer PSentry stores carries three things we needed: the language of the prompt, the assistant that produced it, and the source URLs the answer reported. Sources only exist when the assistant consulted the web, so the first thing we checked was how often that happened. It varied a great deal.

• Claude reported at least one source in 97 answers out of 100, with 12.4 sources per answer on average.
• Perplexity: 94 answers out of 100, 11.6 sources per answer.
• Gemini: 56 out of 100, 10.3 per answer.
• ChatGPT: 44 out of 100, 3.9 per answer.

We then took the domain of every URL and put it in one of three buckets: .com, the national domain of the prompt's language (.it for Italian, .fr for French, .uk for English), and everything else, from .org and .eu to other countries. The rest of this piece reads those buckets language by language.

English: .com is the default source domain

In English, AI answers cite .com sites by default. Across the four assistants, 75 sources out of 100 were .com. Claude went furthest, at 80 in 100. ChatGPT was the lowest, at 68.

The national domain barely registered. Sources on a .uk domain ran from 3.5 per 100 on Claude to 8 per 100 on Gemini, about 5 in 100 overall.

We cannot tell from the data which country an English prompt had in mind. English is the language of international trade as much as the language of the United Kingdom, and a buying question with no country in it is answered from the international web. That web lives on .com.

Country domain or .com in Italian and French

Switch the prompt to Italian and the picture changes. About 50 sources in 100 were .it, against 36 on .com. Counting every EU country domain rather than .it alone, the local share reached 52 in 100.

Three of the four assistants agreed closely. ChatGPT drew 59 sources in 100 from .it, Gemini 55, Claude 55. Perplexity was the exception, with 37 on .it and 43 on .com. It was also the assistant that cast the widest net: 4,944 distinct domains in Italian answers, against 2,437 for ChatGPT.

French sat in between. The .com share stayed in front, at 54 sources in 100, and .fr took about 30. Perplexity again leaned furthest from the national domain, at 23 .fr sources per 100, while ChatGPT and Claude both stayed around 36 or 37.

The pattern is consistent enough to state plainly. The language of the prompt pulls AI sources towards national domains, and how hard it pulls depends on the language and on the assistant.

Spanish and German: too few answers, or none

Two of the five languages PSentry measures are missing from the figures above, for different reasons. German has no data at all: not one of the 17,719 answers came from a German prompt. Spanish has 259 answers, about 65 per assistant. That is enough to look at and not enough to conclude from, so we are not publishing a Spanish split. Anyone selling into Germany or Spain has to run this count on their own prompts.

What the numbers do not say

The analysis comes with four limits, and they matter as much as the figures.

• The prompts are not a sample of the web. They are the prompts PSentry customers monitor, mostly about B2B services and industrial suppliers. Consumer questions could look quite different.
• Language and market travel together. In our data an Italian prompt is almost always about the Italian market. We cannot separate the effect of the language from the effect of asking about Italy.
• A domain is not a nationality. Plenty of Italian and French firms run their main site on .com, and a .it can belong to the Italian branch of a multinational. The count is about domains, not about who owns them.
• These are API answers. PSentry collects them through each assistant's API. The consumer apps can behave differently, and we have not measured by how much.

What a British exporter can take from it

The useful finding is the asymmetry. In English, a .com site sits in the main pool of sources. In Italian, most of the pool is .it. In French it is still mostly .com, with a national third that Perplexity uses less than the others.

None of this means that registering a .it will get you cited in Italy. A domain is a label. The assistant cites pages that answer the question in the buyer's language, and in Italy many of those pages happen to sit on .it. What the count does support is a narrower habit: when you measure a market, look at which domains the answers drew on, not only at whether your name appeared.

Three checks follow from it.

• For each market, list the domains cited in answers to your buying questions, and count how many are national, how many .com and how many something else.
• Check whether your own pages in that language are among them, and on which domain they live. Your French page under a .com and your distributor's site on .fr are two different candidates.
• Note the local sites that keep coming back: trade press, industry associations, regional directories. They show where suppliers get discussed in that market.

The sources view in PSentry lists the domains each assistant used for the prompts you monitor, language by language, refreshed with the two scans a month. It shows where answers come from. It does not change them.

If you have not yet settled which languages to measure, that decision comes before this one. And if the domains you find are mostly review platforms rather than national sites, there is a reason for that as well.