AI chatbots matched bank rates in 77.1% of assessable answers
Survey of 95 regional banks found Gemini had the highest match rate at 82.6%, while 71% of mismatches were unfavourable to customers.
Japanese app development and marketing support company AIRIDGE has published the results of its "AI Search Actual Survey 2026: Regional Banks Edition," covering 95 regional and second-tier regional banks across Japan. The survey was conducted on September 16 and 17, 2026, via API with web search enabled, and asked a total of 5,607 questions to ChatGPT, Gemini and Perplexity.
The survey compared the three AI services on three axes: whether they could read the bank's website, whether they could answer with correct interest rates, and whether they would recommend the bank. For variable-rate housing loans and one-year fixed-term deposit rates, 77.1% of the 1,600 answers that could be assessed matched the figures published on the banks' official websites. Broken down by product, the match rate was 79.3% for housing loans and 75.2% for fixed-term deposits. By AI service, Gemini had the highest match rate at 82.6%.
Of the 366 mismatched answers, 71% were errors in a direction unfavourable to users. About 90% of the mismatched answers cited official websites as their source while giving different figures.
Website structuring and recommendations
AI crawlers could read the websites of 92 banks. Of those, 22 had structured data (JSON-LD) and 8 had structured FAQ data. For 37 banks, no interest rate figures were found in the text of their main product pages.
In response to 522 questions of the form "Where should I consult about inheritance in [city]?", regional banks were recommended in 0 answers.
Quick answers
How accurate were AI chatbots at answering bank interest rate questions?
Of the 1,600 answers that could be assessed, 77.1% matched the figures published on the banks' official websites.
Which AI service had the highest match rate?
Gemini had the highest match rate at 82.6%.
When was the survey conducted?
The survey was conducted on September 16 and 17, 2026, via API with web search enabled.