When Customers Ask for Their Bank, AI Recommends Someone Else
AIVO Meridian's study of 2,160 AI decision conversations finds one fintech recommended outright more often than 30 US banks combined. Banks keep credit and very little else.
A small-business owner asks an AI assistant where to open an account for a new LLC. Across 2,160 decision conversations run on 18 September 2026 on ChatGPT, Gemini and Perplexity, the answer was most often a fintech. Mercury received 383 outright recommendations across the 1,440 ChatGPT and Gemini conversations. All 30 banks tested, combined, received 299. On Perplexity, which searches the web live, the count was 211 to 153. Across all three assistants, fintechs collected 1,346 outright recommendations and banks collected 452.
The study, reported today in American Banker, measures recommendation at the moment an assistant is pressed to make a choice. Awareness research asks whether a model knows a bank. This study asks whether the model tells a customer to use it, and the two questions produce very different answers.
Banks own credit and little else
When a small business names a bank and asks about a credit-led need such as SBA lending or a credit line, some bank takes the final, outright recommendation 74% of the time. The bank the customer actually named wins outright 41% of the time on ChatGPT and Gemini, and 54% on Perplexity.
In the other three situations tested (forming a new LLC, switching providers, and international banking), some bank is recommended outright only 2 to 27% of the time on ChatGPT and Gemini, and 2 to 14% on Perplexity. The rest goes to fintechs. Across these scenarios, the assistants effectively sort the market by product. Banks are treated as lenders, and a business owner looking for anything other than a loan is sent elsewhere.
Being named barely helps
Across the four largest US banks, Chase, Bank of America, Citi and Wells Fargo, being the institution the customer explicitly asked about produced an outright recommendation only 6% of the time on ChatGPT and Gemini. Citi was never recommended outright in 36 test conversations. Bank of America and Wells Fargo were never recommended outright on ChatGPT or Gemini at all.
The turn-by-turn data shows how this happens. When a customer names a bank, the assistant's first answer centers on that bank 65% of the time. By the final recommendation, that share has fallen to 24%. Of the banks that led at turn one, 72% are gone by the decision, and 86% of the time a fintech has replaced them. Unprompted conversations show the same erosion: Chase makes the shortlist in 75% of them and is the outright choice in 35%. Bank of America falls from a 52% shortlist rate to 1% outright.
For money-center banks, the cost goes beyond lost share. A customer asks an assistant about their own bank by name, and the assistant recommends a competitor to their face.
The citation contest
Perplexity cited 30,424 sources across its answers. Fintech companies' own websites accounted for 32% of them, the largest single category. Banks' own sites accounted for 20%, and chase.com was the only bank domain among the 15 most-cited sources. Asked how to open the account, the assistants pointed customers to mercury.com 701 times, more than to the sites of Chase, Bank of America, Wells Fargo and Citi combined (402).
This is the mechanism behind the headline number. A searching assistant retrieves and cites whichever pages it finds most useful for the question, and fintech content currently answers these questions in ways the assistants retrieve more frequently. A bank's published content has become part of its competitive infrastructure. If the assistant cannot retrieve a clear answer from the bank's own site, it will retrieve someone else's.
Verified specialization survives
One result shows the position is winnable. With no bank named, Chase was still the top recommendation for credit-led questions on ChatGPT (73 of 90 conversations) and Perplexity (42 of 90), the only bank to lead any unprompted situation on any assistant.
The top of the named-bank win-rate table follows the same pattern. Huntington and Synovus were each recommended outright in the credit scenario 9 of 9 times, and in 11 of Huntington's 13 outright wins the assistant cited its standing as a top SBA lender. That standing checks out against SBA's own 7(a) data, where Huntington ranked #2 nationally in FY2025 after seven straight years at #1. Synovus, now part of Pinnacle Financial Partners, and SouthState hold the same kind of standing at regional scale. Without any geographic cue in the prompts, the assistants consistently favored institutions whose documented lending strengths matched the scenario.
There is no single AI to win
In 18 bank-and-situation pairs, one assistant recommended the named bank in every repeat while another recommended it in none. Consistency within a single assistant is also low: asked the same question three times, an assistant gave the identical final recommendation in only 43% of cases (307 of 720). A board report that quotes one blended "AI recommendation" figure is concealing this volatility. Measurement has to be assistant-specific, with retrieval conditions kept separate.
What this means for bank leadership
AI assistants now function as a recommendation channel of their own, separate from search, advertising and brand awareness, and they assign reputations product by product. Credit strength does not carry over into deposits, switching or international banking, so those products need their own positioning. A bank's own content needs to answer, clearly and specifically, the exact questions customers are putting to assistants. Every recommendation the assistant redirects represents a customer the bank may never have the opportunity to acquire through its own channels. The institutions that start measuring AI recommendation by product and by assistant, the way they already manage search and social, will be the ones positioned to change these numbers.
About the study
AIVO ran 2,160 directed conversations on 18 September 2026, testing 30 US banks across four small-business scenarios on ChatGPT, Gemini and Perplexity, with three repeats of each. Each was a four-turn conversation, so the results capture a decision reached over a dialogue, which single-prompt testing cannot show. ChatGPT and Gemini answered without web search. Perplexity searched live, and results from the two retrieval conditions are reported separately throughout.
Each final recommendation was classified as outright, shared, routed or no pick under AIVO's published AI Win-Rate Methodology v2.0 (DOI 10.5281/zenodo.22663407).
Percentages are reported only where the 95% confidence range falls within ±20 points. A blind human review of 150 answers agreed with the automated outright-recommendation scoring in 99% of cases. Full bank scorecards and the win-rate table are available on request.
Comments ()