A single test run on September 7, 2026 produced a 60-point swing in airline recommendations depending only on the language of the query. Claude Sonnet 5, tested with web search disabled, recommended British Airways in 52.5% of Persian queries for a London–Tenerife return flight. In Russian, it recommended easyJet 72.5% of the time and British Airways not once. Same model. Same question. Same afternoon.

The dataset behind that result contains 1,650 responses, gathered through DataForSEO's LLM Responses API across six shopping scenarios, five languages, and 40 runs per combination. The tester asked Claude to name three brands and pick one recommendation in English, Arabic, Russian, Simplified Chinese, and Persian. Comparison rounds added ChatGPT (gpt-5.6-sol) and Gemini (gemini-3.5-flash), plus web-search-enabled runs.

These are tool-signal findings from one afternoon, not a confirmed platform statement. Keep that distinction in mind throughout.

Does the language of a query change which brand AI recommends?

The short answer is yes, and sharply in travel categories. In hotels, Bahía del Duque won 43.6% of Chinese responses but only 7.7% of Arabic ones. Hotel Botánico took nearly a third of English recommendations and vanished entirely from Russian responses.

Four other categories proved more stable. Roborock won robot vacuum recommendations in all five languages, at between 87.5% and 97.5%. Zoho CRM dominated its category everywhere, peaking at 95% in Persian. Monday.com led project management software in every language, though ChatGPT broke that pattern — Wrike won four of five languages there, including 86.7% in Chinese.

The pattern suggests categories with a short list of global leaders resist language effects. Categories without consensus — airlines, hotels — do not. That explanation remains a guess by the tester, not a verified mechanism.

What did Anthropic's own research find?

The catalyst for the tests was a July Anthropic report on Claude's values across languages. Anthropic sampled just under 310,000 real conversations across the 20 most common languages, split equally between Sonnet 4.6, Opus 4.6, and Opus 4.7. The company found Claude leaned toward greater warmth in Hindi and Arabic, and greater rigor in English and Russian.

Anthropic did not test whether language changes which brands the model mentions. That gap motivated the independent tests. The tester's own caveat: Anthropic's study covers values, not brand selection, so the two findings are related but not equivalent.

What changes when web search is switched on?

Turning web search on changed the winning brand in seven of ten language-and-category combinations. In the hotel category, the winner changed in all five languages. Ten new hotel names appeared, and previous recommendations disappeared.

One result stands out. With web search off, Claude recommended an Iberostar Grand hotel twice in 198 responses. With web search on, JOIA by Iberostar — the name Iberostar adopted in a July 2024 rebrand — became the most recommended hotel, taking 31% of picks overall and 60% in English. The model did more than swap the new name for the old one. Something it found online pushed that property to the top.

The tester flags a real weakness here: 42.7% of web-search responses hit the output limit before finishing, including 83% in Persian and 63% in Arabic. Brand picks were still captured, but the numbers may be soft.

How reliable is any of this?

The tester states the limits plainly. The sample is small. One afternoon. Six categories. Five languages. Three models. Hotel percentages in particular may sit within the margin of error. One mention in 40 responses is not evidence.

Statcounter's August 2026 figures frame the stakes: ChatGPT drove roughly three-quarters of AI chatbot referrals in Saudi Arabia, Kazakhstan, and Singapore, and the three U.S. models together exceeded 90% in each market. More than half of Los Angeles County residents speak a language other than English at home, including roughly 74,000 Persian speakers.

What should marketers monitor next?

The web-search results carry the practical implication. Training data is fixed and beyond reach. Live web results are not. The findings suggest, though they do not prove, that publishing regional content can influence future recommendations in each language.

Monitor whether these numbers hold week to week, whether the language gaps widen as models update, and whether web-search-enabled answers keep diverging from memory-only answers in your own category.