SwitchWize Research Desk: Exploratory Pilot
Advertising vs. Rate: An Exploratory Pilot
Among 7 institutions with unanimous 3-pass classifier agreement, the two banks whose homepage marketing leads with service (Chase, Wells Fargo) carry the largest live rate gap against the top tracked savings rate (4.12 points each), while the one bank whose marketing leads with a rate number (UFB Direct) carries the smallest (0.87 points). This is a small, disclosed pilot, not a statistically validated finding.
Last reviewed August 26, 2026 · SwitchWize Research Desk
What this is, and isn't
This started as a design for a validated index across roughly 50 institutions with a 150-item human-labeled gold set, following the method in Chen, Hu & Ma's "Banks' Images: Evidence from Advertising Videos" (NBER Working Paper 34220, 2026). Collecting and classifying video advertising at that scale, and producing a genuine human-adjudicated label set, was not something this research pass could responsibly deliver. Publishing labels described as human-adjudicated without an actual human labeling them would be exactly the kind of fabrication this research program exists to prevent.
What follows instead is a small, fully disclosed pilot: homepage hero marketing copy only (no video), for the institutions where a homepage could actually be captured (11 of 20 attempted -- several major sites block automated fetching), classified by three independent LLM passes rather than human raters, with the real agreement rate reported rather than assumed. Every excluded institution and every failed capture is listed below, not just the clean results.
The pattern, in the institutions that classified cleanly
The academic literature (Chen, Hu & Ma, 2026) found that banks with high local market share lead with service and trust while price-focused entrants lead with rate. In this small sample, the two highest-market-share incumbents that classified cleanly (Chase, Wells Fargo) did lead with service, and both carry the largest rate gaps in the sample. But two digital-first challengers (Ally, SoFi) that the paper's framework would predict to lead with price instead led with trust and lifestyle appeal, only UFB Direct matched the "new entrant leads with price" prediction. That divergence is reported as found, not smoothed over to fit the paper's pattern.
| Institution | Position | Marketing lead | Own APY | Gap vs. top rate |
|---|---|---|---|---|
| Chase | Incumbent | Service-led | 0.01% | 4.12pp |
| Wells Fargo | Incumbent | Service-led | 0.01% | 4.12pp |
| U.S. Bank | Incumbent | Trust / emotional | 0.05% | 4.08pp |
| Bank of America | Incumbent | Trust / emotional | n/a | n/a |
| Ally Bank | Challenger | Trust / emotional | 3.00% | 1.13pp |
| SoFi | Challenger | Trust / emotional | 3.10% | 1.03pp |
| UFB Direct | Challenger | Price-led | 3.26% | 0.87pp |
Gap is each institution's own tracked HYSA APY versus the 4.13% top-tracked-average APY used in the Bank Gap Index's August 2026 release, for direct consistency with that published figure. Bank of America classified unanimously as trust/emotional but has no HYSA rate_observations row at research time, so its gap is not computed or guessed.
Classifier reliability, reported as measured
Three independent, blind LLM classification passes were run over the same 11 captured homepages using the identical published rubric below. Rater 2 and rater 3 agreed on all 11 institutions. Rater 1 disagreed with both on 3 of 11 (Capital One, Discover, Fidelity), each a case involving ambiguous or off-target copy. Average pairwise agreement across all three raters: 81.8%. Seven of 11 institutions had unanimous 3-of-3 agreement and make up the clean index above; the other 4 disagreed on category, were flagged as topically off-target by a majority of raters, or both, and are listed as excluded, not silently dropped.
Published rubric
PRICE: Copy leads with a specific rate, APY, fee, discount, or other concrete financial benefit number as the primary hook.
SERVICE: Copy leads with functional convenience, ease of use, customer support, features, or how-it-works framing as the primary hook, without a headline rate/fee number.
TRUST_EMOTIONAL: Copy leads with brand reputation, awards, community, relationships, aspirational lifestyle imagery, or emotional appeal as the primary hook, without a headline rate/fee number and without functional/service framing as the lead.
Excluded from the clean index
Capital One: Captured homepage hero copy was about credit cards ("Take charge of your credit"), not savings/banking. Flagged not topically relevant by 1 of 3 raters, and the captured page itself is a poor proxy for Capital One's savings marketing specifically.
Discover: Flagged not topically relevant by majority (2 of 3) raters -- generic brand tagline capture, ambiguous whether it represents savings-specific marketing.
Navy Federal Credit Union: Captured homepage hero copy was about auto loans, not savings. Flagged not topically relevant unanimously.
Fidelity: Flagged not topically relevant by majority (2 of 3) raters -- generic investing tagline, not savings-specific.
Institutions that could not be captured
9 of 20 institutions initially attempted could not be captured at all -- several major bank sites actively block simple automated fetchers or render their marketing copy entirely in client-side JavaScript. This is itself a real finding about what a reproducible, low-cost pipeline can and can't capture without more sophisticated browser automation.
citi: Homepage is JavaScript-rendered; static fetch returned only navigation chrome, no hero copy.
cit-bank: Automated fetch blocked (HTTP 403).
synchrony: Connection reset during fetch.
american-express: Connection reset during fetch.
usaa: Automated fetch blocked (HTTP 403).
alliant: Fetch timeout.
schwab: Automated fetch triggered an authorization/bot-check error page, not real content.
pnc: Fetch timeout.
marcus: Connection refused.
What a real version of this would still need
- A genuine human-labeled gold set, ideally with a real second rater independent of SwitchWize, to replace the LLM-ensemble reliability signal used here.
- Browser-automation-based capture (not simple fetches) to reach the institutions blocked in this pass, and a real path to TV/video-ad content if that remains the goal, respecting each platform's terms of service.
- A sample large enough (dozens of institutions per category, not single digits) to support an actual correlation estimate with a confidence interval, which this pilot deliberately does not attempt.
How to cite this
SwitchWize, "Advertising-vs-Rate Pilot." Retrieved from https://www.switchwize.com/research/advertising-rate-index.
View the full dataset and methodologyEducational information, not personalized financial advice. Rates are illustrative of current market conditions and should be confirmed with the provider.