Business bank accounts: what AI recommends, Citewake Edition 2 (collected 29–30 Oct 2026)

2Where the AIs disagree

Among the top 15 of each panel, the brands with the largest gap between the AI that recommended them most and the AI that recommended them least.

United States

Europe — composite research panel

3How each AI answers

United States

Gemini answered 11 of 30 questions without running any web search. That is the model Citewake asked, gemini-3.1-flash-lite, through its API with search available. Across all four AIs, 386 searches were run; 23.6% of them included a year. Every page Gemini reported was one it cited, so its pages found and cited match.

Europe — composite research panel

Gemini answered 11 of 30 questions without running any web search. That is the model Citewake asked, gemini-3.1-flash-lite, through its API with search available. Across all four AIs, 361 searches were run; 20.5% of them included a year. Every page Gemini reported was one it cited, so its pages found and cited match.

5Agreement

United States

On 22 of the 30 questions all four AIs recommended a brand first; they named the same brand first on 5.

Europe — composite research panel

On 22 of the 30 questions all four AIs recommended a brand first; they named the same brand first on 8.

6Method in brief

• Questions. 60 scored questions in three tiers (head, mid, tail), written as a buyer would ask them and never naming a brand, in two panels of 30 (United States) and 30 (Europe — composite research panel). See the list. A further 15 validation questions are kept private.
• Asking. Each question was asked once of each AI through its official API, with its own web search on and no location sent. Models: ChatGPT: gpt-6-luna; Claude: claude-haiku-5-5; Gemini: gemini-3.1-flash-lite; Perplexity: perplexity/sonar.
• Collection. 29 October 2026, 13:44–14:06 UTC. 240 valid answers of 240 planned: 204 named at least one brand, 36 named none.
• Reading. One grand master model (Claude Opus 5.5) read the four answers to each question side by side and marked each brand as recommended, mentioned or negative. A brand counts only if its name appears word for word in the answer.
• Measure. Recommendation rate is the share of decision answers in which the AI recommended the brand. The overall rate weights each AI and each tier equally, so it can differ slightly from a raw count. Intervals are 90%, from a cluster bootstrap that resamples whole questions. Each panel is measured on its own questions: United States, 80 decision answers (20 decision questions, 20 answers from each AI); Europe — composite research panel, 84 decision answers (21 decision questions, 21 answers from each AI). 2 popularity questions (which brand is most used, not which to choose) are reported apart and never counted in the recommendation rate.
• Method. v0.9, benchmark v1, scoring run 20261030T054636318Z. The full method: citewake.com/methodology.

7Limitations

• One run. Each question was asked once of each AI, in one 21-minute window. AI answers vary from run to run; the intervals cover variation across questions, not repeat runs.
• Two panels, never pooled. The United States panel's questions name the United States. The Europe — composite research panel panel's questions name Austria, Belgium, Germany, Spain, France, the United Kingdom, Ireland, Italy, the Netherlands and Portugal, or Europe as a whole. Questions were in English and no location was sent, so each panel describes the market its questions name, not where an AI was asked from. The panels ask different questions about different markets; they are never pooled, and a difference between them is not a finding about either.
• Spellings merged. allicabank was counted as Allica Bank; vividmoney was counted as Vivid; liveoakbank was counted as Live Oak Bank; annamoney was counted as ANNA Money.
• Benchmark figures, not market share. These are benchmark visibility figures across a fixed bank of questions with equal weight per tier, not estimates of market share.
• Top picks are rare. Most answers do not name a single best choice. Even Mercury (United States) was the top pick in only 9.9% and Wise Business (Europe — composite research panel) was the top pick in only 2.6% of decision answers, so this report ranks by recommendation rate.

8Disclosure

Citewake is operated by the team that also builds aipozi, Biotree and livedot. Biotree and livedot appear in these categories. They were measured by the same method as every other brand, with no adjustment. None of them was named in this category.

9The questions

Each question links to its page, with the four answers in full, the brands read from them and the pages they cited.

10Data

• Every answer: all 240 in full, with the searches each AI ran, the sites it found and cited, and the brands readCSVJSON
• Every brand in each panel (United States: 166; Europe — composite research panel: 169), every measure, ranked within its panelCSVJSON
• The reportPDF

Licence: CC BY 4.0. Use and share with credit to Citewake.

11How to cite

Citewake. (2026). Business bank accounts: what AI recommends, Citewake Edition 2 (collected 29–30 Oct 2026) (Edition 2) [Data set]. https://doi.org/10.5281/zenodo.23792150

@techreport{citewake-ed2-business-bank-accounts,
  author      = {{Citewake}},
  title       = {{Business bank accounts: what AI recommends, Citewake Edition 2 (collected 29–30 Oct 2026)}},
  institution = {Citewake},
  type        = {Edition 2, data set},
  year        = {2026},
  month       = oct,
  version     = {1},
  doi         = {10.5281/zenodo.23792150},
  url         = {https://citewake.com/editions/2/business-bank-accounts},
  note        = {Licensed CC BY 4.0}
}