Research
/
Engine landscape

AI Answers Dropped the Fine Print on 1 in 4 Fintech Claims That Had It

Published

360 ChatGPT and Google AI Mode answers about 60 public fintechs, November 16 to 17, 2026, with 3,447 claims checked against their cited pages.

Evidence
AI prompt data
Coverage
3,447 claims, ChatGPT and AI Mode
Summarize with AI

ChatGPT and Google AI Mode left the fine print out of 28% of fintech claims whose cited page had it. They kept the number and lost the asterisk. In October, two AI Mode answers gave Mercury's deposit insurance limit without its conditions, and this test measures how common that is. For fintech marketing and compliance leads, a citation is not proof, because 1 in 5 claims were on no page the answer cited.

The short version

  • 1 in 4 claims lost the fine print: 28% of fintech claims whose cited page attached a condition left it out (214 of 776).
  • 1 in 5 claims were on no cited page: 19% of claims in answers with a readable cited page (635 of 3,382).
  • Few claims contradicted the page: 1.6% of claims a cited page stated (43 of 2,747).
  • Prices lost the most fine print: 6 in 10 claims that lost a condition were about price (130 of 214).

What this does not show: That the answers were wrong. A number without its condition is incomplete, not false. A claim missing from its cited pages can still be true.

What to do with this

  1. If a pricing page lists fees, ask ChatGPT and AI Mode the main price questions and look for the fine print. Plan, volume tier, extra fees, and region went missing most, even beside the price.
  2. If a team tracks AI answers about its company, read each cited page for the claim itself. A citation is not a quote.
  3. If deposit protection, safeguarding, or the card issuer sits only in legal documents, state each in plain text on one public page. In answers to protection questions, 28% of claims were on no cited page (228 of 818).

Read more: AI Grounding

Answers left out the fine print on 1 in 4 claims that had it

Revolut's business savings page lists the Scale plan's 2.65% rate. Beside it sits a £1,000,000 balance limit. One ChatGPT answer gave the rate without the limit.

ChatGPT dropped the fine print on about 1 in 4 such claims (86 of 351) and AI Mode on about 3 in 10 (128 of 425), a gap too small to call.

1 in 4 claims lost the page's fine print

Claims whose cited page attached fine print (a condition) to the fact, by whether the answer kept all of it.

Both engines
72%28%
ChatGPT
76%24%
Google AI Mode
70%30%
Kept every conditionLeft out a condition
EngineClaims with a page conditionKept every conditionLeft out a condition95% interval, left out
Both engines776562 (72.4%)214 (27.6%)21.8% to 34.3%
ChatGPT351265 (75.5%)86 (24.5%)17.6% to 32.4%
Google AI Mode425297 (69.9%)128 (30.1%)23.0% to 38.7%
% of claims whose cited page attached a condition · 60 public fintechs · November 16 to 17, 2026 · Omnius test of 360 AI answers checked against the pages they cited

Our read: We think answers keep the headline number and lose the line beside it. It works like a quote with the footnote cut off. Every word is accurate, and the reader still gets a different deal.

We picked this measure after the data came in, and it is the least certain one: two readers of the same 54 answers put it at 24% and 28%.

1 in 5 claims were on no page the answer cited

One ChatGPT answer said Rapyd keeps UK customer funds separate from its own money. It cited Rapyd's Singapore terms, which cover Singapore only.

AI Mode left 23% of its claims unsupported (415 of 1,840) and ChatGPT 14% (220 of 1,542). The gap held across the 60 companies.

Protection and terms answers went off the page

Share of claims that no page the answer cited states, by the question the answer responded to.

Protection questions85 of 382 vs 143 of 436
22.3%32.8%
Terms questions79 of 555 vs 196 of 591
14.2%33.2%
Price questions56 of 605 vs 76 of 813
9.3%9.3%
0%10%20%30%40%
ChatGPTGoogle AI Mode
QuestionChatGPTGoogle AI ModeBoth engines95% interval, both
Protection questions85 of 382 (22.3%)143 of 436 (32.8%)228 of 818 (27.9%)23.1% to 32.7%
Terms questions79 of 555 (14.2%)196 of 591 (33.2%)275 of 1,146 (24.0%)19.8% to 28.5%
Price questions56 of 605 (9.3%)76 of 813 (9.3%)132 of 1,418 (9.3%)6.7% to 12.6%
All questions220 of 1,542 (14.3%)415 of 1,840 (22.6%)635 of 3,382 (18.8%)16.2% to 21.7%
% of claims in answers with a readable cited page · 60 public fintechs · November 16 to 17, 2026 · Omnius test of 360 AI answers checked against the pages they cited

Our read: We think answers fill in protection details from general knowledge, then attach whatever page they found. The claim may still be true. A finance team that clicks through will not find it there.

Counting only answers whose cited pages all loaded, the share falls to 16% (366 of 2,329).

6 in 10 claims that lost the fine print were about price

Checkout.com's pricing page offers free processing to registered charities "in the countries in which we currently operate". One AI Mode answer gave the free processing and dropped the country limit.

Price claims led because they carried most of the fine print (449 of 776 claims). Once a page attached a condition, answers dropped it at similar rates for price, protection, and terms questions.

Plan, volume and extra fees went missing most

Kinds of fine print the cited page attached and the answer left out, across 214 claims.

Plan or eligibilityplan tier, account type, approval
59
Amount or volumethresholds, tiers, balance limits
57
Extra fees or exclusionsfees on failed attempts, excluded methods
44
Regioncountry, currency, domestic only
32
Time or promotioncut-off times, limited-time offers
27
"From" or "up to"a range given as a flat figure
14
Partner bank or pass-throughcoverage through partner banks
7
0204060
Condition left outClaims
Plan or eligibility59
Amount or volume57
Extra fees or exclusions44
Region32
Time or promotion27
"From" or "up to"14
Partner bank or pass-through7
Claims with any condition left out214
claims with each kind of condition left out · a claim can lose more than one · 60 public fintechs · November 16 to 17, 2026 · Omnius test of 360 AI answers checked against the pages they cited

Our read: The fine print lost most often is what decides who gets the price, such as the plan, the volume tier, or the fees on top. Printing it beside the price did not stop answers from dropping it, as the Revolut page shows. That's why we treat a cited number as half the answer until its fine print comes with it.

Methodology

Data. 720 answers from ChatGPT and Google AI Mode, set to the US, collected through Atomic AGI, Omnius's product. Each engine got 3 questions (price, protection, terms) about each of 60 public fintechs, in two runs about 25 minutes apart on November 16 and 17, 2026. Model versions were not logged. We checked one answer per question and engine: 360 answers and 3,447 claims. A cited page stated 2,747 of them, no cited page stated 635, and 65 had no readable page.

Method. Each claim about the company's own price, protection, or terms was checked against the pages its answer cited. Intervals resample the 60 companies. A second reader of 54 answers agreed on 95.7% of claims.

Limits. AI models did the checking, against the cited pages not against the truth. A read of 40 dropped conditions found 3 miscoded.

Sources

Omnius articles this note builds on

Monthly Growth OpenLetter.

Learn how to scale user acquisition without scaling costs from our findings. We spent years exploring, so you don't have to.

Your submission has been received!              
Oops! Something went wrong while submitting the form.

Related research.

Become the answer of ChatGPT Google Claude Gemini Perplexity

Omnius is an AI-native SEO & GEO agency maximizing the probability of B2B SaaS & Fintech companies being included in AI answers, selected by agents, and chosen by humans.

Trusted by
Our technology is used by
Referenced by