Methodology: How We Measured What Households Leave on the Table

Last updated: October 2026

This page documents the study behind How much the average household leaves on the table. It exists because the study’s headline number — a median of $1,376 a year left on the table — is only meaningful next to a clear statement of what was modeled, what was assumed, and what was left out.

The one thing to hold onto: the households are modeled, not surveyed. They are generated by a program, not collected from people. We did not interview anyone, and no row in the dataset describes a real family. What the study measures is the distance between a flat single-card rewards strategy and an optimized multi-card wallet across a realistic spread of income and spending.

What the study does

For each of 1,000 modeled households, we computed two things over the same spending:

1. The optimized wallet — what our optimizer recommends, holding up to four personal cards, with no cards already in the household’s name.
2. A flat single-card status quo — one no-fee card paying a flat rate on every purchase.

The difference is the gap. The article’s headline is the median gap against a 1.5% flat card.

Generating the households

The generator is deterministic. It draws from one seeded random number generator, so the same code and the same catalog produce byte-identical output files, forever.

Income

Household income is a lognormal draw with a median of $87,460 and a shape parameter (sigma) of 0.75, truncated to $12,000–$650,000 and rounded to the nearest $100. The median is the U.S. median household income for 2025 (Census Bureau, released September 2026).

The median anchors the draw to the published U.S. household income distribution; sigma sets the spread. The realized sample has a median of about $86,350 and a mean of about $117,800, which tracks the shape of the published distribution without claiming to reproduce it. The point of the draw is not to match a specific table — it is to make sure we are not modeling 1,000 households that all look like the calculator’s default preset.

Household type

Each household is one of three types: single (30%), couple (34%) or family with children (36%). The type multiplies the category mix, because a one-adult household and a family with children spend differently at the same income:

• Single: less groceries, less utilities, more dining per dollar.
• Couple: close to the base mix.
• Family: substantially more groceries, more utilities, more phone, more “everything else”.

Category spending

Card-eligible spending starts from the calculator’s own Average household preset (the same numbers behind the preset buttons on the rewards calculator) and is scaled by income:

scale = clamp((income / 87,460) ^ 0.55, 0.40, 2.80)

The exponent of 0.55 encodes the well-documented fact that spending rises with income but less than proportionally: doubling income does not double household consumption. The clamps stop the tails from producing implausible budgets.

Each category budget is then multiplied by an independent lognormal draw with sigma 0.18, so two households on identical income still differ. Amounts are rounded to whole dollars.

Deliberate structure

Two features exist because they change the answer, not for realism theater:

• Non-travelers (45%). Nearly half of households are set up with zero airfare and hotel spending. If every household traveled, travel cards would win by default and the study would not be informative.
• Category dispersion. A household that is heavy in one category and light in another is a different optimization problem from a household that is light everywhere. The jitter produces that spread.

What is excluded, and why

How the modeled households compare with real data

The households are synthetic, but they are not arbitrary. Two checks against published government data:

• Income. The modeled sample has a median income of about $86,350 and a mean of about $117,800. The Census Bureau puts U.S. median household income at $87,460 and the mean at $126,700 (2025, released September 2026). The survey reports a mean income before taxes of $104,207 across 135.8 million consumer units.
• Spending. The modeled median household puts $2,366 a month ($28,386 a year) on cards. Summing the BLS survey’s average across the categories most households can pay by card — food at home ($6,224) and food away from home ($3,945), gasoline ($2,411), airline fares ($682), other lodging ($1,347), entertainment ($3,609), telephone services ($1,460), internet ($698), household furnishings ($2,414), apparel ($2,001), drugs ($658), personal care ($978), household operations ($1,921), alcohol ($643) and reading ($125) — gives roughly $26,400 to $29,100 a year, or about $2,200 to $2,430 a month. The modeled median sits inside that range.

The modeled mean spend is higher than the survey’s mean, because spending scales with the income draw and that draw has a long right tail. Treat the study as representative in shape, not in level.

Sources: U.S. Census Bureau, Income in the United States: 2025 and U.S. Bureau of Labor Statistics, Consumer Expenditure Survey 2024, Table 2500.

Finding the optimized wallet

The optimizer is the same code the live calculator runs. Since October 2026 it is an exact search: a branch-and-bound over the eligible cards that returns the same wallet as scoring every combination of up to four cards, under the calculator’s own rules (Year 2+ value, spending split across the wallet within each card’s caps, pooling, and its tie-break toward fewer cards). An earlier, greedy version of the optimizer sometimes missed the best wallet; this study was run with the exact search, and an automated test checks the search against brute force.

“Up to N cards” takes the best of the optimizer’s one-to-N picks, so a larger allowance never scores lower.

Valuing a wallet

Both sides of the gap use the same valuation engine — the one the live calculator uses.

• Horizon: Year 2+. Welcome bonuses are excluded. A signup bonus is a one-time event that would flatter any wallet, and comparing wallets by their bonuses is a different question from comparing them by their ongoing earn.
• Rewards value. Points and miles are valued at The Points Guy’s October 2026 monthly valuations (published 2026-10-01), the same table the calculator uses. Cash back is worth its face value.
• Pooling. Some no-fee cards earn transferable points that are only worth transferring when another card in the same program unlocks transfers. Those cards are valued at 1 cent alone, and at their program’s transfer value when the wallet contains the unlocking card. This is why a wallet can gain value from a free card.
• Annual fees. Subtracted.
• Statement credits. Added at the calculator’s defaults — the credits a typical user is modeled to redeem without stretching. This is the most important assumption in the study, and it is why every gap is reported twice: once with credits, once without.

The credits caveat, in numbers

At the median, the four-card optimized wallet counts about $825 a year in statement credits. Rewards minus fees barely change between a one-card wallet ($974) and a four-card wallet ($972). Most of the difference between a small wallet and a large one is credits, not earn rates.

So the study reports:

• Gap (credits included): median $1,376 against a 1.5% flat card.
• Gap (rewards only): median $551 against a 1.5% flat card, and $405 against a 2% card.

If you will genuinely use the credits, the first number is the right one. If you won’t, the second is.

How this compares with our how-many-cards study

Our how many credit cards should you have study uses the same optimizer, the same valuations and the same assumptions: perfect use (every bill paid in full, the best cards held, the right card on every purchase), personal cards only, Year 2+ value with statement credits counted, no welcome bonuses, rent and spending abroad excluded. The two studies differ in which spending they model, and that explains the apparent conflicts between their headline numbers.

• Bigger numbers in the how-many-cards study. Its grid gives a $10,000-a-month profile the same weight as a $1,000 one, so its overall medians sit at a higher spending level than a typical household’s. Read at matching spending the two agree: at $2,500 a month that study’s second card adds $387 and its third $83, against $371 and $86 here.
• The same shape. Both find the second card does most of the work, the third adds under $100 for most profiles, and the fourth adds little. Both pick the American Express Gold with the Capital One Venture X as the most common pair, and both count a median of $825 in statement credits in the four-card wallet.
• Different questions. The how-many-cards study measures what each extra card adds. This study measures the gap between an optimized wallet and one flat-rate card, so its headline compares against a 1.5% card rather than against a smaller wallet.

The flat baselines

The status quo is deliberately unglamorous: one no-fee card, used for every purchase, paying a flat rate with no categories and no credits.

No signup bonus, no category bonuses, no credits, no annual fee.

The choice of 1.5% as the headline is not arbitrary. The CFPB’s 2025 Consumer Credit Card Market Report puts the rewards actually earned on general-purpose rewards cards at about 1.6% of purchase volume (up from 1.4% in 2021). A flat 1.5% card is therefore a fair stand-in for the real status quo, and the 2% comparison is the generous version of it.

What the baseline is measuring

The baseline is a strategy, not a card count. The average U.S. consumer already has about 3.7 credit cards in active use (Experian, June 2025), and roughly 90% have at least one. The gap in this study is not the value of owning more cards — it is the value of using the cards you already own well. A household with four cards can still put every dollar on one of them, and that is precisely what the flat baseline models: one card, one rate, no decisions.

Reproducing the numbers

Everything is in the repository and reproducible offline:

go run ./scripts/cmd/household-study

The command reads the same cards.json catalog and the same point valuations as the live calculator, loads the catalog into a temporary SQLite file (it never touches the production database), and writes:

• research/household-study/results.json — every aggregate in the study
• research/household-study/households.csv — all 1,000 households, row by row
• research/household-study/by-income-band.csv — the median tables
• static/study/household-study/ — public copies of the same files, linked from the article

The run takes under a minute, needs no network access, and is deterministic: the same catalog and the same fixed seed produce identical bytes. The seed is recorded in results.json under assumptions.seed.

Data dictionary: households.csv

The per-household file has one row per household. The columns worth knowing:

Limitations

1. Modeled, not surveyed. The most important limitation. These are synthetic households built to span a realistic range; they are not a sample of the U.S. population and the study cannot tell you what any real household earns or loses.
2. Point valuations are estimates. TPG’s transfer valuations are a widely used benchmark, not a price. Points are worth less in cash terms, and less still if you never transfer them.
3. Credits are assumed used. The default credit treatment is generous; the rewards-only numbers exist to bound it.
4. Approval is assumed. The optimizer does not model credit score, issuer rules, ⁵⁄₂₄ status or the value of a hard inquiry. Real households cannot always get the cards the optimizer likes.
5. Paying in full is assumed. Rewards are a rounding error next to credit card interest. The Federal Reserve’s 2025 survey of household finances found that 45% of cardholders carried a balance at least once in the previous 12 months. If you are in that group, nothing here is worth acting on until the balance is gone.
6. The catalog is bounded. 95 cards, 64 eligible for recommendations, only 11 that pay cash back. The cash-back-only results rest on a small set, and a card missing from the catalog is a card the optimizer cannot pick.
7. No gaming or manufactured spending. The study assumes spending happens naturally and that caps, rotating categories and activation requirements are handled as the optimizer models them.
8. One seed, one sample. The results are deterministic, which is good for verification, but it also means the study reflects one particular draw of 1,000 households rather than an average over many draws.

Verification

The numbers quoted in the article are tied out against results.json by an automated test in the repository, so an article edit that contradicts the data fails the build. The embedded dataset has its own integrity tests: the household count, the wallet tables for one through four cards, the cash-back view, the income bands and the household types must all be present and internally consistent.

If you find an error, the data is public and the method is above — please tell us. Corrections are more useful than the study.