Ask Claude in Russian and it picked easyJet for a London-to-Tenerife flight 72.5% of the time. Ask in Persian and British Airways led, at 52.5%. In a test of AI shopping recommendations, the language of the question sometimes changed which brand came out on top.
Anthropic Watch analysis
What happened
Reza Moaiandin of SALT.agency tested Claude Sonnet 5 across six shopping and booking scenarios in English, Arabic, Russian, Simplified Chinese and Persian. The main test produced 1,200 responses, with 40 runs for each scenario-language combination. He also tested two scenarios with web search enabled and compared two scenarios with ChatGPT and Gemini, bringing the total to 1,650 responses.
The biggest differences appeared in flight recommendations: Claude favoured easyJet in 72.5% of Russian responses but 12.5% of Persian ones. British Airways was recommended in 52.5% of Persian responses, but not once in Russian. Other categories were steadier: Roborock dominated the robot-vacuum results across all five languages.
Why it matters
If people use AI to compare products, a brand’s visibility may depend on more than the question and the model. Moaiandin’s results suggest language can change the shortlist, even when the scenario stays the same. Checking a chatbot only in English could miss how it answers customers elsewhere, or multilingual customers at home.
These are results from a defined set of prompts and a limited number of runs, not a map of every shopper’s experience. The airline results are striking, but they do not establish why the recommendations shifted or whether another test would produce the same pattern.
Our read
The useful takeaway is less “AI picks brands differently everywhere” than “test the language your customers actually use”. That is a sensible prompt for marketers, and a reminder for shoppers that one chatbot answer is not a buying guide carved into stone. Forty runs per setting make this a worthwhile signal, not a final verdict.
What to watch
- Whether repeated tests produce similar language-based differences.
- Whether the same patterns appear in other models and product categories.
- How web search and regional models affect the recommendations.
Discussion spark: Should companies test AI brand recommendations in the languages their customers use, or is that too small a sample to guide marketing decisions?
Sources and evidence
- Source update (13 October 2026, 12:00 UTC)
Anthropic Watch is independently operated by WittyWires. It is not affiliated with, endorsed by, or operated by Anthropic.