AI quotes the right annual fee 69% of the time

Grounding, in plain English.When an engine "grounds," it searches the live web and builds its answer from pages anyone can read and change. When it doesn't, it recites what it memorized in training, which for credit cards can be a year or more out of date on fees and bonuses. The existence of citations is the proof the answer is influenceable.

Most people pricing the apply decision are asking ChatGPT. On 78% of card queries it searched the live web; the other engines searched on 82 to 98% of the same queries. When an engine doesn't search, it can't see this quarter's welcome offer or this year's fee increase. How much of this month's gap comes from the new model generation is broken out in the model upgrade section below.

Grounding is a behavior, not a fixed property of a model: it moves with the model version and with whether search is switched on. We re-measure it every month rather than treat any single number as a constant, and the value is the trend. These rates also come through our standardized retrieval harness, one shared search layer across all four engines; on its native consumer search the same model can ground more or less often, so the cross-engine gap and the monthly trend are the durable read, not any single rate.

We gave five buyer personas a full conversation with each engine and recorded the decision each one reached. Across 20 conversations, 55% ended in an apply or switch. The premium-cardholder persona was steered to downgrade in 3 of 4 conversations, the clearest retention signal in the study. With 20 conversations this is a directional read of how the engines diverge, not a population estimate, and the divergence itself is the finding.

The bonus-hunter, who weighs offers against each other, opened a card with 1 of 4 engines (applied with 1, no commitment with 1, skipped with 2). Every conversation that ends without a card is a high-intent applicant the assistant had and did not convert.

1 of 4 name a card, and they name 1 different card; 3 end without a recommendation. The acquisition outcome turns less on the buyer's need than on which assistant they happened to open. Across all 20 conversations, 55% ended in an apply or switch; the rest split between downgraded, closed, skipped and no commitment, directional on a small per-cell sample.

45% of every card mention went to the same five cards. AI named 31 cards in total, so the tail is real, but the top five carry the apply decision. Top issuers by share were Chase, American Express and Capital One.

This is the incumbent-bias risk for a challenger card and the moat for an incumbent. AI leans on a small, training-saturated set of household names. The flip side is the tail: the other 55% of card mentions is spread across 26 other cards, which is exactly the contestable ground for a brand that gets its inputs in front of the engine. The next question is where that bias comes from, because that is where the tail gets won.

cnbc.com was the most-cited source at 10.7% of citations, followed by capitalone.com (6.4%) and americanexpress.com (5.8%). The most-cited issuer page, capitalone.com, took 6.4%, so AI cited cnbc.com about 1.7 times as often as the biggest issuer's own site. Within the 20 most-cited domains, issuer sites took 24.8% of all citations and sites the issuers don't own took 47%; the long tail holds the rest. Editorial and press sites alone took 19.2%.

One caution on this section: these sources come from our shared search layer. When we asked six of this report's questions through ChatGPT on OpenAI's own search in October 2026, it cited about five sources per answer, mostly the issuer's own card pages plus NerdWallet or Bankrate, and mostly recommended the same cards. For ChatGPT itself, your own card pages likely weigh more than this chart shows.

The content behind these citations had a median publish age of about 6 months, and 40% was more than a year old, even though 30% carried a recent "last updated" stamp. Fresh-looking metadata on old facts is how a last-cycle bonus survives into a current recommendation.

58% of the reasons AI gave for a card were about money: the fee, cash back, rewards or the bonus. Those are facts you publish, and they are facts AI often gets wrong.

No annual fee leads at 18% of all apply reasons. The sign-up bonus takes 11%, and fee-aversion (no or low annual fee) adds 22% in total, across 989 reason tags. No single reason dominates the way price dominated streaming: cashback, sign-up bonus and travel rewards follow. The bonus is the sharpest reason for acquisition, and it is the one fact that rotates every quarter.

101 of 146 cited annual fees were accurate against the current Schumer Box. The cards that raised their fee are among the least accurate: the Chase Sapphire Reserve (50%) and the Amex Platinum (53%). AI quoted a pre-increase fee 11 times for the Chase Sapphire Reserve and Amex Platinum alone. A second kind of miss shows up on no-fee cards: AI attached a fee to the Capital One Quicksilver in at least four of every ten mentions.

For the October 2026 edition the measured engines moved to a new model generation. To tell a real category shift from a model change, we re-ran July 2026's models on a frozen subset of 21 of the same credit-card prompts in the same cycle. Held constant, the category barely moved: the only statistically significant change from July 2026 to October 2026 on the old models was Hilton Honors (−8.3 pp). What did move was the models themselves.

The clearest effect is grounding, how often an engine searches the live web before answering. On the same prompts, the upgrade changed it sharply, and in opposite directions by engine.

The upgrade did not narrow the shelf this time. Of the 31 tracked credit-card brands, 15 were named more often by the new models than the old ones on the same prompts, 6 less often, and the rest held. The largest moves were a change of label: American Express +14.3 pp and Amex Platinum −10.7 pp. Both are American Express, so read it as the new models naming the issuer more and the card less, not as a change in how often they recommend it.