The questions we ask, how we run them, what we record, and how each number is worked out. Published so every figure in the report can be checked.
The set holds the fifty highest-intent consumer money questions, spread across the categories that matter to retail banking: cards, deposits, lending, and investing. These are the questions where an assistant names specific banks and makes a recommendation, the moment the report is built to measure. The full question set is published alongside the results.
We put each question to ChatGPT, Claude, Copilot, Gemini, and Perplexity. Every run happens in a fresh session with a US framing, with the question typed in as a consumer would type it and nothing added. This quarter we ran each question once per assistant, so the results read as a directional snapshot of where the answers sit today.
Banks named, in the order they appear. Card names roll up to the issuer, sub-brands roll up to their parent, and name variants of the same company count once, so Capital One 360 counts toward Capital One and PNC Bank counts toward PNC.
Sources cited, as links, footnotes, or grounding references, each reduced to its registered domain.
The top pick, where the answer settles on a single recommendation.
Answer Share. How often a bank is named in the answers for a category.
First-Mention Rate. How often a bank is the first name in the answer.
Citation Share. Among answers that show a source, how often a given website is the one cited.
Owned-Citation Rate. Out of every source cited, how often it is a bank’s own website.
Every number here traces back to a row in the open results file, so you can check it yourself. We label any estimate as an estimate and keep it apart from the measured counts. The raw results ship with the report, and we post any correction in place with the date.
Assistant answers shift with wording, location, account state, and model version. The fixed question set and the clean session hold steady what we can, and we describe the rest here so you can weigh it.
