Mapou Research · Issue 003 · Credit Cards · June 2026
Which credit card AI tells your applicant to open.
First-time applicants are asking AI which card to get, and people with your card are asking whether to keep paying the fee. We put the exact questions they ask, "best card to start with," "is the annual fee worth it," "what's the current sign-up bonus," to ChatGPT, Gemini, Claude, and Grok, then measured which card each one named, the reason it gave, and whether the fee and bonus it quoted were current. 245 measurements across the four engines, plus 20 full buyer conversations.
Three of the findings below should change what your acquisition and retention teams ship this quarter. Each one comes with the specific move it implies, and who owns it. Start at the top.
The same applicant gets a different card from each assistant. Your acquisition mix may depend on which AI someone asks.
We put the same buyer through a full conversation with all four engines. The card they were steered to changed by engine: one excellent-credit traveler was pointed at a $95 card by one assistant and a $795 card by another. Across 20 conversations, 65% ended in an apply or switch, and the premium cardholder was told to downgrade in nearly every conversation. The recommendation path is not stable across assistants, and that path is your top-of-funnel.
WHAT TO DOKnow which engines name your card to your target applicant, and which name a competitor, before you spend another dollar on paid acquisition. The apply decision is being made in the chat window.
Read the full finding ↗AI gets the annual fee right only 60% of the time, and it's worst on the cards that just raised fees.
Across every cited annual fee, 60% were accurate. The misses cluster on the cards that changed price in the last 18 months: the Amex Gold and Chase Sapphire Reserve are still quoted at their old, lower fees. AI quoted a discontinued or pre-increase fee 16 times for the Reserve and Platinum alone. An applicant pricing the apply decision against a stale fee is deciding on bad data.
WHAT TO DOGround-truth the annual fee, APR, and current welcome offer on the surfaces AI reads, then re-measure monthly. The fee is the single fact the apply decision turns on, and it's the one AI is most often wrong about.
Read the full finding ↗The sign-up bonus is the single biggest reason AI gives, 12% of all apply reasons, and bonuses change quarterly.
Sorting every recommendation reason, the welcome offer is the top apply-decision lever at 12%, just ahead of cashback and travel rewards. Combined with fee-aversion (18% no-or-low annual fee), money facts dominate the apply decision. Welcome offers rotate every quarter; the bonus AI quotes is the one most likely to be last cycle's.
WHAT TO DOTreat your current welcome offer as a fact you publish, not just an ad. The pages AI reads should carry today's bonus, minimum spend, and window, because that's the number it leads with.
Read the full finding ↗TWO MORE FOR THE ANALYSTUnderneath those three: this month ChatGPT grounds (searches the live web) on only 26% of card queries, against 84 to 100% for the other three engines (grounding gap), and AI shapes its apply recommendation mostly off creator videos and comparison sites, not your own page, with youtube.com the single most-cited source (where it comes from).
Cadence
Monthly · acquisition-led
Method
50 prompts · 4 engines · 5 personas
Confidence
Wilson 95% on every proportion
Read on
Full report ↘Findings as of June 2026 · refreshed monthly · AI Acquisition Intelligence
How four AI assistants recommend
credit cards.
We ran 50 credit-card queries through ChatGPT, Gemini, Claude, and Grok. Each query is voiced like a real shopper, not a benchmark prompt. We then put five buyer personas through 20 multi-turn conversations, asking each AI to help them apply, keep, close, downgrade, or switch a card.
The study measures recommendation behavior, not search rank. For each engine we record which card was recommended, why it was framed that way, and whether the fee and bonus it quoted were current. 245 measurements in total.
For a card-acquisition team, the question is whether AI names your card to a first-time applicant, with today's welcome offer and annual fee. For retention, the question is whether AI is telling your cardholder to keep, downgrade, or close when they ask "is the fee still worth it." This report measures both, with acquisition as the primary focus.
50 prompts · 4 engines · 5 personas · 245 measurements
Grounding rate by engine
The assistant most applicants use is the least likely to check current terms.
ChatGPT grounding rate on credit-card queries. The other three engines ground far more reliably.
vs 84 to 100% for Gemini, Claude, and Grok · 50 prompts × 4 engines
Most people pricing the apply decision are asking ChatGPT. On 26% of card queries it answered from memory rather than the live web. The other engines searched on 84 to 100%of the same queries. When the engine doesn't search, it can't see this quarter's welcome offer or this year's fee increase.
Measured in June 2026, with each engine on its default web-access setting. 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.
What you can do about it · with mapou
If AI answers your applicant from stale memory, your current bonus and fee never enter the conversation.
mapou measures, per engine and per query type, when AI grounds on your card and when it recites. We map the queries where it answers cold, so you know exactly where the live web has to carry your current terms.
A per-engine grounding map for your card, refreshed monthly, so you see which assistants are answering applicants from training data.
mapou services
Persona × engine decisions
The same applicant, steered to different cards by different engines.
We gave five buyer personas a full conversation with each engine and recorded the decision each one reached. Across 20 conversations, 65% ended in an apply or switch. The premium-cardholder persona was steered to downgrade in nearly every conversation, the clearest retention signal in the study.
| Persona | ChatGPT | Gemini | Claude | Grok |
|---|---|---|---|---|
| Maya, 24 · thin credit file, first card | Applied Discover it Secured | Applied | Applied Discover it Cash Back | Applied Discover it Secured |
| Devon, 31 · travel bonus-hunter, churns cards | Applied Chase Sapphire Preferred | Inconclusive | Skipped | Applied Chase Sapphire Reserve |
| The Ahmeds, late-30s · family spenders, cashback-led | Applied | Kept | Applied Blue Cash Preferred | Applied Costco Anywhere Visa |
| Linda, 52 · premium cardholder weighing renewal | Downgraded Amex Platinum | Downgraded Amex Platinum | Inconclusive | Downgraded Amex Platinum |
| Marcus, 29 · carrying a balance, debt-consolidator | Applied Citi Double Cash | Applied Wells Fargo Reflect | Applied Wells Fargo Reflect | Applied Wells Fargo Reflect |
One applicant · four assistants · the same question
A frequent traveler with excellent credit asks each assistant the same thing: which card is worth opening for the bonus right now. All four say apply. No two name the same card.
ChatGPT
Amex Gold
mid-tier rewards, $325 fee
Gemini
Amex Platinum
premium travel, $895 fee
Claude
Amex Business Platinum
a business card, to a personal applicant
Grok
Chase Sapphire Reserve
premium travel, $795 fee
Four assistants, four cards, four annual fees from $325 to $895, and one of them a business card the applicant never asked for. The acquisition outcome for this buyer is decided less by their need than by which assistant they happened to open. We saw the same fracture on a debt-consolidation applicant: two assistants gave the correct 0% balance-transfer card, one steered to a rewards card that does not solve the problem.
What you can do about it · with mapou
AI is not neutral in the apply conversation. It converges on a short list, and your card is either on it or it isn't.
mapou runs your real buyer personas through the same multi-turn conversations every month, recording which card each engine talks them into, and why. You see the apply decision as the buyer experiences it, not as a static ranking.
A persona-by-engine decision map showing exactly which applicants AI steers toward your card, and which it steers away.
mapou services
Recommendation concentration & where it comes from
AI names the same five cards no matter who is asking.
of every card mention went to just five cards, out of 31 that AI named at all. The recommendation set barely moves as the applicant's profile changes.
card-level RSOV across 199 responses · top issuers Chase, American Express, 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, so a card outside the top five is largely invisible at the apply moment even when it fits the applicant better. The next question is where that bias comes from.
of every source AI cited for a credit card was youtube.com, ahead of any issuer's own page or review site.
1,576 citations · 214 domains · next source nerdwallet.com at 8.9%
Put differently: AI cited youtube.com about as often as it cited every major issuer's own page combined. The place the apply recommendation gets shaped is creator videos and comparison sites, not your application page. The content behind those citations had a median publish age of about 4 months, even though 18% carried a recent "last updated" stamp. Fresh-looking metadata on old facts is how a last-cycle bonus survives into a current recommendation.
What you can do about it · with mapou
The pages that decide whether AI recommends your card are mostly not your pages.
mapou identifies the exact comparison pages, forums, and creator content each engine cites for your card, then checks which carry your current fee, bonus, and approval guidance. You get a ranked list of the inputs that actually move the recommendation.
A source map of the third-party pages shaping AI's view of your card, sorted by how often each engine cites them.
mapou services
Why it matters
In these conversations, AI funneled the apply decision through two facts above all others: what the card costs, and what the bonus is worth. Both are facts you publish, and both are facts AI often gets wrong.
What drives the recommendation
The apply decision runs on money: the fee, then the bonus.
of all apply reasons are the sign-up bonus, the single biggest lever, and the fact that changes most often.
fee-aversion (no or low annual fee) adds another 18% · 931 reason-tags
No single reason dominates the way price dominated streaming. Instead the apply decision spreads across four near-equal money-and-reward levers: no annual fee, cashback, the sign-up bonus, and travel rewards. The bonus is the sharpest of them for acquisition, and it is the one fact that rotates every quarter.
No annual fee
The card costs nothing to hold. AI leads with $0 as the reason to pick it.
Cashback
Flat or category cash back framed as the core value.
Travel rewards
Miles, transfer partners, and portal value for travelers.
Sign-up bonus
The welcome offer drives the recommendation. The single biggest apply-decision lever, and the most volatile fact.
What you can do about it · with mapou
If AI leads with the bonus and the bonus it knows is last quarter's, it is selling your applicant a card that no longer exists on those terms.
mapou tracks which reason each engine leads with for your card, and whether the bonus and fee attached to that reason are current. When the lever is the welcome offer, we check the offer.
A reason-by-engine breakdown for your card, flagged wherever the money fact AI is leading with has gone stale.
mapou services
The catch
AI is accurate on the cards that haven't changed and wrong on the ones that just did. The fee increase you announced last year is the fact it is most likely to miss.
Annual-fee accuracy & the confusion tax
AI quotes the right fee 60% of the time, and misses on the cards that just raised it.
of cited annual fees were accurate against the current Schumer Box. The errors cluster on recently-changed cards.
55 of 92 cited fees within tolerance · annual fee only
Annual-fee accuracy · cards with ≥8 cited fees
| Card | Actual fee | Accurate | Cited |
|---|---|---|---|
| Amex Gold | $325 | 29% | 4/14 |
| Chase Sapphire Reserve | $795 | 53% | 9/17 |
| Capital One Venture X | $395 | 58% | 7/12 |
| Amex Platinum | $895 | 62% | 8/13 |
| Chase Sapphire Preferred | $95 | 80% | 12/15 |
Accuracy = cited fee within the greater of 10% or $5 of the verified current fee. Cards with fewer than 8 cited fees are excluded here; per-card rates with small samples are shown with their denominator so the reader can weigh them.
The confusion tax
7×
AI recommended the Citi Custom Cash (closed 2026-05-28) to a new applicant. It closed to new applicants days before this run, so it can no longer be obtained.
8×
AI quoted the old $550 fee for the Chase Sapphire Reserve, after it was raised.
8×
AI quoted the old $695 fee for the Amex Platinum, after it was raised.
What you can do about it · with mapou
Every time AI quotes a fee that's too low or recommends a card that's gone, it sets an applicant expectation you can't honor at the application.
mapou ground-truths your current fee, APR, and welcome offer, then measures per engine how often each one quotes them correctly. When a number drifts, you see which engine, which page, and how far off.
A monthly fee-and-bonus accuracy score per engine for your card, with the stale sources that caused each miss named.
mapou services
The throughline
You can't change how the model was trained. You don't need to. Every assistant that searches the live web builds its answer from pages you can change, and mapou shows you, per engine, where it grounds and where it still recites from memory.
What you can change
You can't change the training. You can change the pages it reads.
Most of what AI says about your card is downstream of pages that exist on the web today, and that grounded fraction is exactly the part you can move. The loop is the same one SEO taught, one layer up:
- 1
Find where AI looks
Identify the comparison pages, forums, and creator content each engine cites for your card.
- 2
Sort by what you control
Split the inputs into your own pages (application, terms, FAQ), pages you influence (the outlets AI cites), and pages you can only dispute.
- 3
Fix the flagged inputs
Get your current fee, APR, and this quarter's welcome offer onto the surfaces AI actually reads.
- 4
Re-measure on a frozen prompt set
Re-run the same queries next month and prove the fee accuracy and recommendation share moved. The before/after is the product.
What the recommendation channel is worth
AI recommendation is an acquisition channel. Price it the way you price the others.
You already pay a known cost to put your card in front of a comparison-site reader. The assistant answering "which card should I get" is the same shelf, except placement there is earned through the pages AI reads, not bought. Put your numbers in: the only figure we assert is the recommendation share, the lever this study measures and tracks over time.
AI card-shoppers / month
People asking an assistant which card to get. Your number, or an estimate.
Your recommendation share
The % of those answers that name your card. This is the number mapou measures.
Your blended CAC / funded card
What you already pay per acquisition on paid + affiliate. Card CAC typically runs $100 to $250.
2,500
AI-sourced applications a month at your current recommendation share.
$438k
In monthly acquisition value riding on this channel: spend it offsets, or leaks to whichever card AI names instead.
$438k
Added each month if your recommendation share doubles, the lever this study measures and tracks.
Illustrative. Volume and CAC are your inputs, not mapou figures. Recommendation share is the one number we measure and improve. This models acquisition-channel value, not lifetime value or revenue.
The honest limit: a handful of facts are baked into the model's training and only move on the model's clock. A rename or a brand-new card can lag for months no matter what you publish. We measure that lag rather than promise to erase it. Paid AI placement is a complementary lever, and the audit data makes its targeting smart, but the durable moat is the organic-retrieval loop plus the longitudinal dataset.
What to do with this
See how AI frames your card before your applicants do.
The apply decision is moving into the chat window, and right now it's being made on facts you didn't supply and can't see. Three things this study says your team should do this quarter.
Own the fee and the bonus on the pages AI reads. AI quotes your annual fee correctly only 60% of the time, and it leads with the welcome offer more than any other reason. Those are facts you publish. Get this quarter's bonus, minimum spend, and current fee onto the comparison and creator pages AI actually cites, not just your own site.
Win the apply queries, where the high-value decision happens. A new card is worth far more per acquisition than a single subscription, and the applicant asking "best card to start with" is the highest-intent buyer there is. Know which engines name your card to that buyer, and which name a competitor, before you spend another dollar on paid acquisition.
Watch the renewal conversation, because AI is steering it. When the premium cardholder asked whether to keep paying, AI steered them to downgrade in nearly every conversation. That is a retention event happening off your property. Measure what AI tells your cardholders at renewal, monthly, so you see the churn signal before it lands on your statement.
This report is the category view. The version that moves your numbers is scoped to your cards: which applicants AI steers toward them, what fee and bonus it quotes, which third-party pages shape that answer, and how all of it moves month over month once you fix the inputs.
AI Acquisition Intelligence
See how AI frames your card at the apply moment.
The shoppers deciding which card to open are asking an assistant first. We measure how AI frames your card's fit, fee, and offer, surface where those inputs are wrong or stale, and track how the framing moves once they are corrected.
mapou.aiCategory name: Conversational Commerce Intelligence. What this measures: mapou measures how AI assistants alter commercial decisions inside high-intent consumer categories. For credit cards, the buyer-specific wedge is AI Acquisition Intelligence.
How to read these numbers. This study prioritizes behavioral depth over broad population sampling. Findings represent observed recommendation behavior under a fixed prompt set, and are directional unless repeated across re-runs. The value is in the shape of the result and its month-over-month movement, not in any single point estimate.
50 evergreen credit-card prompts × four engines (ChatGPT, Gemini, Claude, Grok), routed through a single retrieval layer, plus 20 multi-turn buyer conversations across five personas. 245 total measurements. Recommendation Share of Voice (RSOV) is the share of responses naming a given card. Every per-engine proportion carries a 95% Wilson confidence interval; small-sample per-card rates are shown with their denominator.
Annual-fee, APR, and welcome-offer ground truth was sourced from each issuer's own pricing page and the Schumer Box on each product's rates-and-fees disclosure, with per-field confidence labels. Annual fee and foreign-transaction fee are the high-confidence accuracy spine; welcome offers are labeled separately because they rotate quarterly and are partly geo-targeted. The study runs monthly; any single capture is a snapshot, and the value is in the month-over-month delta.