ChatGPT drew 90.9% of its cited sources from regulators and government guidance, compared with 17.1% for Gemini Flash-Lite. Half of Gemini Flash-Lite’s sources came from providers’ own websites
200 of the study’s 213 provider mentions came in the six questions asking which provider to choose. Across 96 answers about rules and risks, no provider was named
When the same questions were asked again about two weeks later, only around a third of cited sources appeared in both answers
Advice warnings appeared in 89.6% of answers about choosing a provider, but in 47.9% of answers to the most basic awareness questions
Edinburgh, 19 November 2026. Four AI platforms drew on strikingly different sources when answering the same questions about consolidating UK pensions, according to new research from JOAN UK, an Edinburgh consultancy that helps financial services firms see what AI assistants tell their customers before they reach the website. The sources shown also varied substantially when the questions were repeated.
The study, Where AI Assistants Get Their Answers About UK Pension Consolidation, put 30 questions a consumer might ask to ChatGPT, Gemini Flash-Lite, Google AI Overviews and Google AI Mode. No question named a provider, so every firm in the results was introduced by the AI itself. Each question was asked twice on each platform, producing 240 answers. The questions, what would be measured and nine predictions were written down before any answers were collected.
Four platforms, four source environments. With each question weighted equally, regulators and government guidance bodies such as MoneyHelper, GOV.UK and the FCA made up 40.9% of the source mix, against 4.6% from forums and social media. The platforms differed sharply. On ChatGPT, 90.9% of cited sources were regulatory or government guidance and 4.3% were provider websites. On Gemini Flash-Lite the pattern reversed: 17.1% regulatory and 50.0% from providers. MoneyHelper was the single most-cited source, cited 170 times across 27 of the 30 questions.
Providers appear when the consumer starts choosing. Almost all provider mentions came in the six questions asking which provider is best, where each platform named at least one provider in 10 of its 12 answers. Across the 96 answers covering rules, restrictions and risks, such as transferring a defined benefit pension or avoiding a scam, no pension provider was named.
Cited is not the same as named. Being used as a source and being named often came apart. One provider’s website was drawn on in 41 answers while the provider was named in 10; another was cited in 30 answers and named once. Several investment platforms showed the reverse, named far more often than their own websites were cited.
Sources differed substantially between observations. When each question was asked again about two weeks later, the average share of cited sources that appeared in both answers was 35%, ranging from 28% on Gemini Flash-Lite to 47% on Google AI Overviews. Questions about settled rules were the most consistent; open-ended and buying questions varied most. ChatGPT’s average number of sources shown per answer fell from 4.4 in the first round to 1.6 in the second, with no visible change to the product.
Advice warnings are least common at the start of the journey. The line between guidance and regulated financial advice was flagged in 89.6% of answers to questions about choosing a provider, compared with 47.9% of answers to the earliest awareness questions. This was the reverse of what the researchers had predicted, and the report records it as such. The answers were collected before the FCA, in October, called on AI providers to help consumers understand the limits of the financial information and advice their models generate, through clear and consistent disclaimers.
JOAN UK presented the findings to financial services leaders at a live online session, Is AI Sending Customers Your Way?, on Wednesday 18 November.
Anurag Ahluwalia, co-founder of JOAN UK, said:
“A customer can now ask an AI assistant about combining their pensions long before they reach a provider’s website, and none of that conversation shows up in a firm’s own analytics. What this research shows is that it isn’t one consistent information environment. Different platforms draw on very different sources, and when we asked the same questions again, most of the cited sources were different.
“For providers, being used as a source and being named are different kinds of visibility. Firms need to understand both, and they need to measure them more than once. Official guidance already anchors many pension answers. The opportunity isn’t to replace those sources, but to be a clear, accurate source that AI can use alongside them.” The full report, the first volume of JOAN UK’s AI Discovery Series, and a recording of the session are available at joandigital.co.uk/sessions/ai-answers-uk-pensions.
