Highlights
- The dataset: The analysis behind this study covers over 220,000 AI source citations from prompts specific to the business banking industry, with AI responses from four AI models run in 12 countries during Q2 2026.
- YouTube slid from #1 to #10: Losing roughly 65% of its citation share between the two halves.
- Reddit is the single most-cited domain overall: About 4% of all citations, climbing to #1 in the back half of the window.
- 23% of ChatGPT's citations are from user-generated domains: Reddit and YouTube drive by far the highest user-generated share of any model.
- 58% of Google AI Overview's citations are from commercial domains: The most brand-direct model, it will quote the bank's own page.
- 74% of top sources differ between any two models: Only 26% average overlap of top cited domains, swinging from 18% to 48% by pair.
Which sources should you target to get cited as a business banking brand?
Each model has its own distinctive top sources and citation character:
| Model | Top sources | Citation character |
|---|---|---|
| ChatGPT | reddit.com, wise.com, forbes.com, reuters.com, finder.com | Community + news-led |
| Google AI Overview | youtube.com, airwallex.com, wise.com, qonto.com, venn.ca | Brand-direct + video |
| Grok | reddit.com, youtube.com, wise.com, forbes.com, nerdwallet.com | Social + money-transfer specialists |
| Microsoft Copilot | rankia.com, forbes.com, businessexpert.co.uk, kreditio.se, venn.ca | Regional portals + editorial comparison |
A business-banking brand optimizing for one AI model is absent from roughly 74% of the domains the other models prefer. There is no single AI search result to win, there are four.
How have AI source rankings changed over time in the business banking industry?
Splitting the ~15.5-week window at its midpoint and comparing each domain's citation share in the first half against the second surfaces large movements among the top sources.
The headline reversal is video versus community. YouTube fell from the #1 most-cited domain to #10, shedding about 65% of its citation share, while Reddit rose to #1, growing its share by roughly 140%. Several regional and specialist players climbed hard: businessexpert.co.uk jumped from #25 to #9 (the period's biggest climber), and rankia.mx (+10 places), bbva.com.ar (+8), and the Canadian fintech venn.ca (+6, into the top 5) all gained ground. On the way down, kreditio.se, boekhouder.nl, and bbva.com each dropped 7–8 places.
What type of content do AI models cite for business banks?
We classified every cited domain by category, commercial, editorial, user-generated, reference, institutional. Commercial sites (banks, neobanks, fintechs, and money-transfer providers citing themselves or each other) dominate every model, but the mix beyond that is where the models diverge.
Google AI Overview is the most brand-direct, sending 57.7% of its citations to commercial domains, it will quote the bank's own page. Microsoft Copilot is the most editorial, with 37.8% going to comparison and review media (and the highest reference share at 8.4%). ChatGPT is the crowd model: 23.4% of its citations go to user-generated content like Reddit and YouTube, more than double Microsoft Copilot's UGC share. A small uncategorized residual of 3.5%–6.7% per model is not shown.
For a business bank, the optimization target changes per model: own-site content and product pages move Google AI Overview, third-party comparison reviews move Microsoft Copilot, and community presence on Reddit moves ChatGPT.
Do AI models cite local-language content for business banks?
Local-country domains take about half of citations, and never lead when the question is in English. In the English-language markets, national ccTLDs like .co.uk and .com.au take roughly a third of citations in the UK, Canada, and Australia (33–38%), and essentially none in the US, where sources sit almost entirely on generic .com domains. For non-English markets we use a country-code top-level-domain (ccTLD) proxy, and across all non-English prompts only 51.2% of citations went to a local ccTLD.
Swedish and Dutch business-banking questions stay overwhelmingly local; Spanish questions are pulled toward .com finance portals and away from any single country's domain, unsurprising given the Spanish prompts span Spain, Mexico, Argentina, and El Salvador, fragmenting the "local" signal across four ccTLDs. This is a ccTLD proxy, not true language detection, so these figures understate genuine local-language coverage and should be read as a directional floor.
Which AI model relies most on local sources for business banks?
The localization gap between models is real and consistent. Reading the same six languages model by model, Google AI Overview is the heaviest user of local sources in most markets, while Grok most often defaults to global English-language and .com domains, most visibly in Spanish, where only 30.7% of Grok's citations carry a local ccTLD.
Two patterns stand out. In Dutch and Swedish, every model leans heavily on local sources, Microsoft Copilot reaches 89.6% local in Swedish and Google AI Overview reaches 86.3% in Dutch. In Spanish, every model cites local sources sparingly, with Grok and Microsoft Copilot below 35%. Google AI Overview returned no cited sources for French-language business-banking prompts in this window, so that cell is blank rather than zero.
Should business banks optimize for each AI model separately?
Yes, almost entirely. We built each model's most-cited domains for business banking and compared every pair to see how many sources they share. The average overlap is 26.4%, meaning 73.6% of top-cited sources differ between any two models.
Google AI Overview and Grok are relative cousins, they share 13 of their top domains (48.1% overlap), both leaning on youtube.com, wise.com, Airwallex, Qonto, and Reddit. But Microsoft Copilot is the outsider: it overlaps just 17.6% with both Grok and Google AI Overview, preferring regional finance portals (rankia.com, businessexpert.co.uk, kreditio.se) and editorial comparison content the others largely ignore.
How many sources does each AI model cite per answer?
The models differ by more than 4x in how many sources they pull into a single answer.
Grok cites 26.9 sources per response, about 4.2x ChatGPT's 6.4, and a single Grok answer reached as many as 132 cited sources. Grok's spread is also enormous (standard deviation ~26), reflecting a model that sometimes dumps dozens of citations and sometimes only a handful; the other three are far more consistent. For a brand, Grok offers many more slots per answer but each carries less weight, while ChatGPT's short citation lists make every included source disproportionately valuable.
Context
This analysis draws from Temso's AI visibility monitoring platform, which tracks how brands appear in AI model responses across ChatGPT, Microsoft Copilot, Grok, and Google AI Overview. The dataset covers business-banking prompts, choosing, comparing, and recommending business bank accounts, neobanks, and fintech alternatives, across 12 countries and 6 languages between 11 March and 29 June 2026.
It measures which sources AI models cite, not whether those sources are accurate or whether the recommendations are good. A citation is not an endorsement. These findings reflect citation behavior for business-banking prompts in the monitored markets, which skew toward English-speaking and Western European/Latin American regions, and may not generalize to all institutions or prompt types.
Methodology
How we measured this
We tracked four AI models responding to business-banking prompts across 12 countries and 6 languages over roughly 15.5 weeks. Each response was parsed to extract source citations: the URLs, domains, and metadata referenced. Domain categories (commercial, editorial, UGC, reference, institutional) were assigned from a global domain registry. Cross-model overlap was measured by comparing each model pair's top-20 most-cited domains. Temporal analysis split the observation period at its midpoint and compared domain rankings, by citation share, in each half. Model coverage varied across the observation window, so temporal movements are reported as relative ranks within each half.
The analysis covers over 220,000 cited sources from 22,683 AI responses. Sample sizes exceeded recommended thresholds for every reported finding. Localization is measured by a country-code-domain proxy (the source records carry no detected content language), so those figures are directional and exclude English-language prompts, which sit overwhelmingly on generic domains.
Frequently asked questions
Do different AI models cite different sources for the same business-banking question?
Yes, dramatically so. Any two models share only about 26% of their top-20 cited domains, meaning roughly 74% of the sources one model relies on are absent from another's top list. There is no single AI search result to optimize for.
Which model should a business bank or neobank prioritize?
It depends on the content you can produce. Google AI Overview rewards your own commercial pages (57.7% of its citations are commercial). Microsoft Copilot rewards third-party comparison and review media (37.8% editorial). ChatGPT rewards community presence, 23.4% of its citations go to user-generated content like Reddit and YouTube.
Does writing in the local language help business-banking AI visibility?
In non-English markets, often, but less than you might expect. Across non-English prompts, only about 51% of citations went to a local-country domain. Swedish and Dutch questions stay overwhelmingly local (~73%), while Spanish questions skew toward global .com finance portals (~35%). Among English-language markets, the US pulls essentially no local-ccTLD sources, while the UK, Canada, and Australia sit near a third.
Which AI model relies most on local sources for business banks?
Google AI Overview and Microsoft Copilot rely on local sources the most in most European markets, Microsoft Copilot hits 89.6% local citations in Swedish and Google AI Overview hits 86.3% in Dutch. Grok relies on local sources least in Spanish (30.7%), defaulting more often to English-language and .com domains.
How many sources does each model cite per answer?
Grok cites the most by far, about 27 sources per response, with some answers exceeding 130. ChatGPT cites the fewest at about 6. That 4.2x gap means each source in a ChatGPT answer carries far more weight than one of Grok's many citations.
Are AI citation rankings stable over time?
No. Over ~15.5 weeks, YouTube fell from the most-cited domain to #10 while Reddit rose to #1, and a UK finance portal climbed 16 places into the top 10. AI source rankings churn fast, so visibility has to be defended continuously, not won once.
Why does Reddit appear so prominent for business banking?
Reddit is the single most-cited domain overall and is especially central to ChatGPT and Grok, which lean on community discussion when recommending business accounts. For business-banking brands, that makes authentic, well-regarded Reddit presence a genuine AI-visibility lever, not just a community-management afterthought.
Can these numbers tell me whether a recommendation is accurate or trustworthy?
No. This study measures which sources models cite, not whether those sources are correct or whether the recommendation is sound. A citation reflects what a model surfaced, not an endorsement of quality.
