What AI assistants tell prospects and existing customers about your credit card today, and what you can change about it. Fifty open US consumer cards, read across five assistants, scored on the answer a customer is reading right now.
Assistants describe US credit cards to customers every day. The numbers below say how well they do it, and where a bank can change the outcome.
Fifty cards on one scale, measured on answers customers are reading today.
What is published today is doing the sorting. That is the lever with the shortest handle.
The median card is stated correctly 86 percent of the time and credited to the bank 45 percent of the time.
Citi Double Cash leads all fifty cards. Capital One holds the strongest issuer average at 71.3.
If a customer asked an assistant about your flagship card today, do you know what it says?
It picks the facts, the conditions, and the source. The bank page is often the second thing the customer reads.
Customers ask which card suits them, what it earns, and what it costs. An assistant answers all of it in one pass, in its own words.
It reads what it can find on rates, rewards, and conditions, then names a source for the claim. That source is usually a third party.
By the time the customer opens it, the comparison has been made and the shortlist is set.
Publish the facts an assistant needs before the customer asks for them. Write the card page as evidence, not as a brochure.
AI Quality spans 20.6 points across the fifty cards. Trust Upside spans 9.1. The answer already in the market creates 2.3 times more separation than the upside from clearer evidence.
Correct the facts, conditions, and sources on the page before commissioning another market test.
AI Quality measures the answer customers read now: are the facts correct, current, complete, and credited to the bank. Trust Upside estimates how much clearer evidence could move trust, consideration, verification, and intended action. The first is observed. The second is modelled. Source: Trust Transfer Index analysis, August 2026. n = 50 cards.
The line below is the whole field: fifty open US consumer cards on one scale. The strongest sits at 75.3, the weakest at 54.7. Everything a customer reads about the other forty-eight falls between them.
A twenty point spread on a measure customers act on is a real distance. It separates a card that explains itself from a card that leaves the explaining to someone else.
Find your own cards in this field before the next planning cycle. A card you have read an AI answer for is a card you are managing.
Each card is scored on whether the facts are correct, current, complete, and credited to the bank across five assistants and three question types.
The dot is each issuer's portfolio median. The line is the distance from its lowest to its highest scoring card. Capital One sits at the top with the tightest span in the study: a customer gets the same quality of answer whichever of its cards they ask about.
A high median with a short line is the strongest position on this chart. It says the standard behind the best card has reached every other card in the house.
Take your clearest card page and make it the house template. A standard travels further than a campaign, and costs less.
Single-card issuers appear as one point. Source: Trust Transfer Index analysis, August 2026.
75.3, the top score in the study. It leads because there is almost nothing to interpret: one promise, stated once, in language an assistant can quote without hedging.
One earning rule, no tiers to summarise, and every footnote points the same way as the headline.
What a customer asks about lives together on the card's own page, current and dated, where it can be read in a single pass.
Unambiguous wording gets repeated as written, and the bank keeps its name on the claim.
Simplicity is a design choice, available to any issuer. Write the next card page so there is nothing left to interpret.
Which of your cards could a customer explain back to you correctly after one answer?
The median card scores 86 percent on correct and current facts, and 45 percent on bank source authority. The 41 points between those two numbers is the clearest job on this page.
Review sites and news outlets receive the credit while the bank's own facts are accurate. The customer reads a correct sentence with a third party's name on it.
Make the official facts easier to find, easier to pull, and easier to quote. One dated fact page per card, written to be read by a machine.
41 points of source authority sit inside the bank's own control, and the content behind them is already written.
Your facts are right and a third party is quoted for them. What is that worth, and to whom?
Fact accuracy holds near the 86 percent median on all five. What differs is whether the bank's own page is shown as the source. ChatGPT and Claude do it far more often than Gemini and Copilot.
Perplexity cites constantly and names the bank some of the time. Source: Trust Transfer Index analysis, August 2026. Claim-level diagnostics, ranked.
What works on one assistant carries partway to the next. An assistant that shows your page is the one sending the customer to it.
Test each assistant separately, and read the answers yourself once a month.
Do you know which assistant your best customers use, and how it describes you to them?
Trust Upside estimates how a customer's response could change if the evidence were clearer. Each bar is the average share of available points earned across the fifty cards. The bars near the bottom hold the most headroom.
Median base score across the fifty cards.
Lowest base score in the field.
Highest base score in the field.
A directional estimate from a synthetic panel rather than a market result. Use it to choose which cards to test first, then measure real customer behaviour.
Citi Double Cash leads all fifty cards because its terms are simple, stated in one place, and easy to quote. Any card can be written that way.
One fact page carrying the earning rules, the redemption value, the transfer rules, and the conditions that change any of them. Dated, in plain sentences, on the bank's own domain.
One page carrying the intro period, the deadline, the fee, how purchases are treated, and two worked payoff examples. An assistant quotes a number that exists as a number.
Same structure, same headings, same date stamp. A portfolio reads consistently to a machine when it is written consistently by people.
Give every card one fact page, written the same way.
Start with a flagship, a growth card, and one you already suspect reads badly. What works there becomes the standard for the rest.
Every assistant, every card in the set, with transcripts in front of the people who write the pages.
One dated fact page per card. Current terms, plain conditions, written so a machine can quote them whole.
Same questions, same assistants, one quarter later. What moved is the measure.
Approve the cards and the standard their pages will be written to.
A quarter from now, which single number do you want to have moved, and by how much?
Read the results as a direction rather than a scoreboard. The specification was locked before collection began, and every figure traces back to a recorded answer.
Fifty open US consumer credit cards, across eight categories and twelve issuer brands.
Five assistants answered three card questions twice, in fresh sessions. Natural discovery added category visibility.
Every card received a current first-party source package before scoring.
The final audit covers 10,923 claims. Independent blind coding and adjudication resolved the determinate differences.
A 5,000 agent synthetic panel produced at least 60,000 seeded assignments across three scenario bands.
AI Quality comes from observed answers. Trust Upside comes from a synthetic experiment and is a directional estimate only.
Are the facts correct, current, complete, and credited to the bank, in the answer running today.
Modelled change in trust, consideration, verification, and intended action if the evidence were clearer.
How often the bank's own page is shown as the source for a claim about its own card.
It can be written as deliberately as the product itself. The Trust Transfer Index, August 2026, runs to 15 pages and is free to open.
Download the report