CardRight tells you which card to use for every purchase by counting what your cards cost you, not just what they pay you. It runs on real cards and real store rules, replays your last year, shows where money leaked, and ranks the moves worth making.
Each card has its own rules: bonus categories, spending caps, quarterly activations, annual fees, foreign fees, statement credits and protections that nobody remembers. So most people pick a favorite card and use it for everything.
Apps already exist to help. CardPointers, MaxRewards and Kudos tell you which card earns the most points at a store. They are useful, and most are built around one question: which card pays the most?
CardRight scores every way to pay on one number: what you actually keep.
What Jordan's "5% card" really earned him last year. He carries a balance on his Discover at 26.49%, so every purchase started paying interest the day he made it. A rewards-first app would keep recommending it.
What the same spending would have kept on the right cards: his no-fee 2% Active Cash for new spending, paid in full, while the balance gets paid down. That is $1,029 left on the table.
| What it considers | Points-first apps (typical) | CardRight |
|---|---|---|
| Best rewards card at a store | Yes | Yes |
| Interest when you carry a balance | Rarely central | Core of the math |
| Caps, activations and sign-up deadlines across the whole wallet | Partly | Planned across the month |
| Tax effects (HSA, business cards, card processing fees) | No | Yes |
| Paid card placements | Often part of the model | None, and it can say "no new card" |
| Needs card numbers or a bank login | Often for full features | Never. Only which cards you hold |
The comparison describes the category's general focus as I understand it from public reviews, not every feature of every app.
Real cards make this concrete. Two of the cards in the demo changed their terms in the four months before I built it, and store rules like Costco's Visa-only policy decide which cards even work. The two jobs that need judgment, reading card terms and recognizing merchants, are AI-shaped. The decision itself is plain, tested arithmetic, so the same purchase always gets the same answer and every answer can be checked.
Card fine print becomes rates, caps, fees and credits. Anything unclear is flagged for a person.
terms.jsMerchant to category. Known stores from a list; unfamiliar ones read by a real model, checked by the name rules. Unsure means it asks.
model.js · merchants.jsMoney kept for every way to pay, with each term shown.
optimize.jsA year of spending rerun, losses split by cause.
replay.jsMoves ranked by dollars, each with its math.
plan.jsRecognizing merchants now runs on a real model (gpt-oss-120b on Groq), through the same small server and key as my Umbra product. The model only names a category from CardRight's fixed list; it never sees a card or a balance and never picks a card. If it is down, CardRight uses its own name rules and says so. Reading card terms is still a transparent rule-based stand-in, and a person would approve every new card's rules before they go live.
Costco warehouses take only Visa credit cards. Her American Express Gold, Freedom Flex and Double Cash are all out, and the app says why before she reaches the register. Her Sapphire Preferred is a Visa, so it wins. It is also $380 from a 75,000 point sign-up bonus, but she is on track to reach that with her usual spending, so CardRight does not pretend this one purchase is worth more than its rewards.
He carries $3,400 on his Discover at 26.49%, so a $55 grocery run on it costs more in interest than any reward pays. CardRight says use the 2% Active Cash and pay it in full, in one sentence.
398 purchases replayed. Most of what he lost was interest, not rewards. The line shows his "kept" going below zero while the right cards stay positive.
HSA money goes in before tax, so covering a crown from it saves about 24% at his tax rate. That saving is the same whichever card he swipes, so CardRight tells him to pay with his best rewards card and reimburse himself from the HSA, and counts both in the same currency: dollars kept.
The IRS-approved processor charges 1.75% for credit cards. His everyday card earns 1.5%. On a $3,000 estimated payment that is a loss, so the answer is a free bank transfer.
Route spending by category, turn on the quarterly 5%, and fee checks that are honest both ways: her $325 Gold card only just earns its fee, while Sapphire Preferred gets a prompt to downgrade before its fee renews. Her sign-up bonus is checked too: she is on track without changing anything, so it is not counted in the yearly total.
Rates, caps, credits and bonus progress for each real card. Choose whether you use points for cash or travel, or flip "I carry a balance on this card", and every recommendation updates.
Published statistics, the checkout rules, each card's terms and the pages they came from, plus a plain list of what is assumed.
| Person | Spent | Actually kept | Could keep | Left on the table | Biggest cause |
|---|
| Step | Tuned set | Held-out set | Held-out misses flagged | Misses not flagged |
|---|
The rules alone were the weak link: only about half of unfamiliar names landed in the right category, though every miss came back marked "not sure", so the app asked instead of guessing. The model, run on the same 18 held-out names, got 17 right. Its one miss, Midstate Gas Company, it called a gas station and marked sure; it is a utility bill. That is the dangerous kind of mistake in money, so the app keeps the name rules as a second opinion: when they are confident and disagree with the model, it asks the person. With both together: 17 of 18 right, and no silent mistakes. Answers came back in about a third of a second.
The engine needs which cards you hold and a few facts about them. Linking accounts would sharpen it, but it is optional.
The "next card" check runs market cards against your real spending and can, and usually does, say no new card is worth it.
Rewards, interest, fees, tax and estimates are separate lines. Every card carries the date it was checked and the pages it was checked against. What is assumed is listed as assumed.
Reading terms and recognizing merchants can be wrong, so both report confidence and flag doubt. The recommendation is deterministic.
It explains trade-offs with the math shown, like which balance costs most. It does not tell anyone how to invest or which loan to take.
A small subscription, priced against what the replay shows a person leaves on the table. Or licensed to banks and credit unions as a feature inside their own app, where helping customers use the bank's cards well improves retention.
Dollars kept per user against their own replay baseline. Share of checkouts where the recommended card was used. Interest avoided for people who carry balances. How often the plan's top move gets done within 30 days.
People with three or more cards, and especially people who carry a balance. They have the most to gain and are the least served by points-first tools.
Any recommendation that sends new spending to a card with a carried balance. Any stale card term older than its review window. Any sign that a paid relationship influenced a result.
The mark is a payment card whose stripe turns into a check mark: the right card, confirmed. In the wordmark, a small gold card dots the i in Right. Forest and cream feel like money handled calmly, and gold is reserved for the money you keep.
