The Compressed Market

An analysis step resolved each institution’s brand, aliases, official domains, and competitor set from its public website. The audit then posed a battery of unbranded, industry-standard retail-banking questions to each platform through its production API, with web-augmented answering where the platform provides it. Every prompt carries the suffix “in United States” — a deliberate design choice. An assistant does not natively know where a logged-out user is located, and when no location is given it tends to default to global digital players (Revolut, Monzo, bunq, Wise) and to mix banks from different countries in a single answer. Pinning the geography forces every answer onto the same market frame, so results are comparable across institutions and relevant to the US market being measured. For every answer, the pipeline recorded the full text, the brands mentioned, whether the audited institution was among them, and every source the platform cited.

The 28 prompt frames in the study — identical for every institution, each answered by all three platforms (roughly 246 answers per frame):

The published corpus applies four scope rules. First, the study covers ChatGPT, Claude, and Gemini. Second, the index covers US-chartered, US-focused institutions; audits of foreign-parent US operations were excluded (TD Bank was retained as a judgment call: its US retail subsidiary operates as a major domestic bank). Third, credit unions are out of scope for this edition, which covers commercial and retail banks. Fourth, answers that failed prompt validation were removed, along with four extraction artifacts stripped from mention lists.

Brand names were then canonicalized: roughly 45 alias families were folded so that “Ally”, “Ally Bank”, and “Ally Financial” count as one brand, “Chase” and “JPMorgan Chase” as one, and so on. This step materially changes the results — before folding, Chase’s dominance was split across two names and Ally’s across three.

The cleaned corpus: 6,881 answers (ChatGPT 2,289, Claude 2,296, Gemini 2,296), 82 institutions each with a uniform sample of 84 answers (28 frames × 3 platforms), and 8,519 canonical brand-mention events.

Two metrics carry the paper. Unbranded visibilityis the share of an institution’s answers that mention it. Share of voiceis a brand’s share of all canonical mention events in the corpus.

One property of the measurement basis matters for interpretation. Brand mentions are detected against a tracked panelper audit — the audited institution plus its resolved competitor set, aliases included. Because the major national and digital brands appear in nearly every audit’s panel, cross-corpus share of voice measures them reliably; brands outside every panel are not counted, so the corpus understates the long tail. Every concentration figure in this paper is therefore a conservative reading: adding untracked long-tail mentions could only widen the set of names, and spot checks of raw answer texts confirm that the additional names are dominated by the same national and digital-first brands that already lead the tracked results. A systematic full-text scan of the most prominent untracked brands — the global neobanks — is reported in §3.6.

Because the battery is identical for every institution, per-institution estimates share a uniform sample size (84 answers), and no statistical floor is needed; corpus-level claims use all 6,881 answers.

Across 6,881 answers to 28 differently framed questions, the tracked answer space resolves to 61 brands, and the concentration within it is steep:

Of the 82 institutions, 15 were ever mentioned in an unbranded answer and 67 (82%) sit at exactly zeroacross their 84 answers each. The zero set is not composed of obscure institutions: it includes M&T Bank, KeyBank, First Citizens Bank, Comerica, Valley Bank, East West Bank, UMB Bank, Hancock Whitney, and Prosperity Bank.

The zero-visibility result cannot be explained by ignorance. In a named-recognition control — each institution asked about directly, by name — the assistants recognized and accurately described them in 99.1% of cases, a figure that holds from money-center banks to community institutions. The assistants hold detailed knowledge of Fulton Bank and Hancock Whitney; they simply never volunteer either. What the data shows is a gap between an institution’s presence in the model’s knowledge and its presence in the model’s recommendations — and for most of the market that gap spans the full distance from 99% to zero.

An objection worth pre-empting: perhaps this is one vendor’s quirk. It is not. The three platforms were measured independently, and they produced substantially the same ranking:

Classifying every mentioned brand by institutional footprint shows how the compressed answer space is allocated. National retail banks take 72.5% of all mentions and digital or direct banks 23.3%; regional and super-regional banks combined receive 1.9%. On the subject side, every seeded community and regional bank sits at zero; among super-regionals only Regions, Huntington, and Fifth Third register, each in single digits.

Global digital players such as Revolut, Wise, and Monzo sit outside most audits’ tracked panels, so §3.1’s share-of-voice figures do not count them. To measure them directly, we ran a full-text scan of the answer corpus. The result confirms both their presence and its limits: roughly 6% of answers name at least one global neobank, and the set is narrow — Wise in 5.7% of answers, Revolut in 1.9%, Nubank in a handful; Monzo, N26, bunq, and Starling never appear at all.

Two properties define how they show up. First, they are specialists, not challengers to the shortlist: over 80% of their appearances are concentrated in two prompt frames — international transfers and expat banking — exactly the use cases where their product genuinely leads. Second, when they appear, a major US brand appears beside them in 97% of cases. The assistants position Wise or Revolut as a line-item for a specific job, next to Chase or Bank of America for the banking relationship; in almost no answer does a global neobank displace the US shortlist.

This is also the clearest validation of the geographic pin (§2.1). In unlocated queries, assistants readily lead with Revolut, bunq, or Nubank and mix markets in one answer; with the geography fixed to the United States, that tide recedes to a specialist niche and the answer space resolves to the US market actually being measured.

Aggregates conceal which questions each brand owns. The table below shows the three most-mentioned brands per topic, as a share of that topic’s answers. Chase leads fourteen of the twenty-two topics; Ally owns the price-sensitive and online-native topics (lowest fees, easiest to open, switching, online-only); Bank of America owns mobile apps, innovation, expats, and wealth management.

Where do the assistants get these answers? For the two platforms with transparent citations (ChatGPT and Claude; Gemini emits only opaque redirect URLs and is excluded here), the cleaned answers carry 32,364 citations across 388 distinct domains — with heavy concentration at the top:

Within the seeded list, the correlation between an institution’s asset rank and its unbranded visibility is weak (r ≈ −0.28). The clearest illustrations run in both directions: SoFi, eighty-fourth on the asset-ranked list, holds the third-highest unbranded visibility in the corpus, while M&T Bank and KeyBank — both far larger — sit at zero. Whatever produces AI visibility, it is more available to a digitally native marketing operation than to a balance sheet.

The decisive follow-up. For a panel of regional institutions, run the same intents under three framings: national (“best bank for a small business in the United States”), state and metro (“…in Ohio”, “…in Charlotte”), and unlocated (“best bank for a small business”). Every outcome is informative. If assistants localize well, the compressed national shortlist matters only for unlocated queries, and the strategic advice becomes content-driven. If they localize poorly and recommend the national shortlist regardless of stated geography, a much stronger claim about regional erasure becomes defensible with evidence. If they hedge harder under local framing, regionals are absent from the channel under every framing that exists.

A quarterly re-run converts a snapshot into an index: movement becomes measurable, model updates become visible events, and the leaderboard acquires a time axis. The pipeline that produced this corpus is fully automated, so the marginal cost of the time series is low.