When someone asks an AI assistant which credit card to get, the answer usually comes with an annual fee attached. mapou checked those fees against what each issuer publishes, two ways.
First in full conversations: 20 simulated card shoppers talking it through with ChatGPT, Gemini, Claude or Grok until they decided. The assistants quoted 24 annual fees along the way, and 7 were wrong.
Then at scale, in single answers to card questions: 127 fees quoted, 39 wrong. That's 31% off by more than 10% or $5, whichever is larger. Conversation or single question, it comes out at about 3 in 10.
The numbers
Cards quoted 8 or more times. Correct means within 10% or $5 of the fee in the card's Schumer Box (the issuer's rates-and-fees disclosure). 127 fees across all 15 cards.
The premium cards that raised their fee are the worst. The Amex Platinum costs $895 a year. AI got that right 8 times out of 17. The wrong answers included $695 and $795, older prices for the same card. The Chase Sapphire Reserve is $795. AI got it right 8 times out of 16, and quoted $550 among the misses.
No-fee cards get a fee they don't have. The Capital One Quicksilver and Wells Fargo Active Cash both cost nothing to hold. AI attached a fee to each in half of its mentions (4 of 8 and 4 of 8).
Long-standing fees are fine. Cards that have kept the same fee for years, like the Chase Sapphire Preferred and the Citi Double Cash, were right almost 9 times in 10.
What it means
mapou's read: AI is quoting the card's fee history, not its fee. Price changes take a long time to wash out of the pages AI reads, so the cards that changed most recently carry the most wrong numbers. For an issuer that's a real problem in both directions. A fee quoted too low sets an expectation the application can't meet. A fee quoted on a no-fee card talks the shopper out of it.
What this doesn't show
This covers annual fees only, from one month of answers. Some cards were quoted only a handful of times, so their percentages move a lot with one answer. Welcome offers and APRs drift too, but they change often enough that mapou tracks them separately.
How we measured this
- Conversations
- 20 synthetic card-shopper conversations, September 2026: 5 personas (for example a 24-year-old opening a first card and a 52-year-old deciding whether to renew a premium card), each run once against ChatGPT, Gemini, Claude and Grok, until the shopper decides.
- Single answers
- 50 credit card questions asked once of each engine.
- Ground truth
- Each card's annual fee from its Schumer Box, the issuer's official rates-and-fees disclosure, checked September 2026. A quoted fee counts as correct within the larger of 10% or $5.
- Matching
- Every dollar amount in an answer that sits next to fee language is attributed to the nearest card named before it. 'No annual fee' next to a no-fee card counts as a correct $0.
- Engines
- ChatGPT (gpt-5.4-mini), Gemini, Claude and Grok through one shared search layer with web search on.
- Limits
- Annual fees only. Conversation quotes are few (24), so per-card conversation numbers are not shown. Welcome offers and APRs are tracked separately.
Cite this article
mapou, “AI quotes the wrong annual fee on almost a third of credit cards,” mapou.ai, October 4, 2026. https://mapou.ai/articles/ai-credit-card-annual-fees-wrong
Reporters can reach mapou through for more on the method, other breakdowns or a comment.
Is AI quoting your card's fee correctly?
mapou checks what AI assistants say about your products every month, against your own published terms, and shows which sources the wrong numbers come from.
Monthly research newsletter
Get the next article by email.
Short findings like this one, plus the monthly panel report. Unsubscribe in one click.
We never share email addresses. One unsubscribe link in every send.