Mapou Research · Issue 003 · Credit Cards · July 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. 232 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.

Findings as of July 2026 · refreshed monthly · AI Acquisition Intelligence

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. 232 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 · 232 measurements

Most people pricing the apply decision are asking ChatGPT. On 74% of card queries it answered from memory rather than the live web. The other engines searched on 78 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 July 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. 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, 45% 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. With 20 conversations this is a directional read of how the engines diverge, not a population estimate, and the divergence itself is the finding.

"No commitment" is a result, not a gap: the buyer ended the conversation without the assistant moving them to a decision. It clusters on the bonus-hunter, who weighs offers against each other and walks rather than commit to one card. For an acquisition team that is the live signal: a high-intent applicant the assistant had in the conversation and failed to convert.

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 half of recommendation share is spread across two dozen 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.

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 14% carried a recent "last updated" stamp. Fresh-looking metadata on old facts is how a last-cycle bonus survives into a current recommendation.

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.

No single reason dominates the way price dominated streaming. No annual fee leads, but the apply decision still spreads across a cluster of money-and-reward levers: cashback, travel rewards, and the sign-up bonus close behind. The bonus is the sharpest of them for acquisition, and it is the one fact that rotates every quarter.

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.

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.

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. 1

   Find where AI looks

   Identify the comparison pages, forums, and creator content each engine cites for your card.

2. 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. 3

   Fix the flagged inputs

   Get your current fee, APR, and this quarter's welcome offer onto the surfaces AI actually reads.

4. 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.

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.

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.

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.

Methodology v2 · before / after

What changed when the AI models upgraded.

This month the measured engines moved to a new model generation. To tell a real category shift from a model change, we re-ran June 2026's models on a frozen subset of the same credit-card prompts the same day. Held constant, the category barely moved. 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 also made answers more concentrated: 30 of the 37 tracked credit-card brands were named less often by the new models than the old ones on the same prompts. Newer models give shorter, more decisive answers, so a few names hold, and the long tail thins. That is a model property, not a change in the category, and it is worth watching: the shelf AI shows a shopper is getting narrower.

Before/after measured on a frozen 21-prompt subset, old and new models run the same day. Grounding change is the new model minus the old model on identical prompts.