Methodology

Every verdict on this site is computed, not written. This page is the complete formula, the data behind it, and the things it deliberately does not claim.

Sources

Complaints, recalls and investigations come from the NHTSA public APIs. Fuel economy, electric range and annual fuel cost come from EPA fueleconomy.gov. Raw API responses are cached and versioned, and every page links the sources behind its own numbers.

Data vintage and refresh. The ownership dataset is extended and re-checked every night, and the whole site is regenerated from it every night. Each page carries the date it was last built. Complaint and recall counts are cumulative federal records, so an older model year has had more years in which owners could file — the score corrects for that by working in complaints per year of exposure.

Reliability score (0–100)

The score answers one question: compared with what this same nameplate normally attracts, how much trouble did this model year attract? It is built from quantities that do not depend on how many cars were sold, because raw complaint counts do.

Why that matters. A nameplate selling 350,000 cars a year collects more complaints than one selling 20,000, however good it is. An earlier version of this score divided complaints only by years of exposure, and it punished popular cars for being popular — the kind of error that inverts the real ranking. It was replaced with the model below.

score = 100 − within_model − cross_model − severe_recalls − other_recalls

• within_model (max 42) — this model year's complaints per year of exposure, divided by the median rate of the same nameplate across all of its years. Sales volume is roughly constant across a model's years, so dividing by the model's own median cancels volume out. The ratio is shrunk toward 1.0 with weight n / (n + 40), where n is the complaint count: a year with a thin record is judged close to the model's own norm rather than being read as evidence of excellence.
• cross_model (max 12) — the percentile of this car's complaint rate across the whole index, on a log scale. This is the one term sales volume still leaks into, which is exactly why it is capped at 12 points and computed logarithmically.
• severe_recalls (max 18) — 3 points per safety-critical campaign. A recall is "severe" when its NHTSA summary matches fire, crash, injury, stall, brake-failure or steering-loss patterns.
• other_recalls (max 8) — 1 point per remaining campaign. Recall counts are issued per defect, not per car sold, so they are directly comparable between models.

Because every term is bounded, the lowest score the model can produce is 20/100. No car is branded with a number implying a total failure the data cannot support.

Evidence strength

Each verdict carries a confidence label: strong at 150 or more complaints on record, moderate at 50 or more, thin below that. A thin record is an absence of evidence, not evidence of quality, and the shrinkage term above treats it that way.

Verdicts

BUY ≥ 70 · CAUTION 50–69 · AVOID < 50. Where complaint data is unavailable the verdict is DATA PENDING — never guessed. A model year is only given its own page when it has at least 30 complaints or at least 3 recall campaigns on record, or is an EV with published battery data. Everything thinner is folded into the model overview instead of being published as a page of its own.

Running cost

Annual running cost = EPA annual fuel or energy cost + an age-indexed maintenance band drawn from industry averages (AAA "Your Driving Costs", CarMD Vehicle Health Index). The headline figure is the midpoint of that band, and the band itself is always shown beside it. Running cost excludes purchase price, insurance and depreciation, and the pages say so.

Repair costs

repair = flat_rate_hours × local_labour_rate + parts_band. Flat-rate hours are the typical published times for the representative job in each NHTSA component group; the labour reference is $120/hour at a US independent shop; parts are a low–high band for a quality aftermarket or OE part. Every figure is an estimate of what that class of failure costs — never a quote for a specific car, and never a prediction that the repair will be needed. The jobs shown on a page are ranked by the components that owners of that exact model year actually complain about.

Purchase price, depreciation and insurance

NHTSA carries safety and EPA carries fuel economy. Neither carries money, and there is no free, complete, per-model-year price dataset, so this site does not pretend to know what one specific car is worth. What it publishes is a class-level band, computed the same way for every car and shown with the band around it. The constants are published at /assets/price-model.json.

1. Segment. Each model-year is placed in one of fifteen segments by its own record: an explicit nameplate keyword first (an F-150 is a pickup), otherwise EPA combined economy as a size proxy, then the brand tier re-reads the segment upward for a premium or exotic marque.
2. The equivalent new car today. segment_share × the current US average new-vehicle transaction price (Cox Automotive / Kelley Blue Book). A mid-size car is 0.62 of that average, which is where a new mid-size sedan actually lists. Working in today's money is what stops a 2015 car being priced in 2015 dollars against a 2026 used market.
3. Value today. equivalent_new × retained_value(age) × segment_retention × used_market_index. The retained-value curve is the consensus of the published depreciation studies (iSeeCars, Edmunds, NADA); the segment multiplier reflects their spread — pickups and sports cars hold value, large luxury sedans and used EVs do not; the used-market index (currently 1.08) carries the Manheim index's standing gap above its pre-2021 trend, and comes out when that gap does. A running car never falls below a floor.
4. Price when new is the same figure walked back to the model year with the transaction-price series. Where Wikipedia's infobox carries a published list price for the nameplate, that real figure anchors the calculation instead, and the page says so.
5. Depreciation is the difference between the value today and the value five years from now — on a five-year hold this is normally the largest single line, larger than fuel, maintenance and insurance together, which is why it is shown before them.
6. Insurance is the national average annual full-coverage premium for the segment (NAIC / industry averages), relieved by vehicle age because full coverage tracks the value at risk, then re-priced by your country's insurance index.

The published band is ±18% around the central estimate, roughly the interquartile spread of real listings for one model-year in one market. It is not a valuation, a dealer price, a trade-in offer or an insurance quote — condition, mileage, options and location move a real car materially. Every price panel takes the price you are actually being quoted and recomputes the whole picture from it.

Owner satisfaction

Complaint records tell you what broke. They cannot tell you whether the owner would buy the car again, so we ask owners directly, one response per account per car. Averages are published only once five owners have answered — below that a mean is noise wearing a number's clothes. Responses are never edited, never weighted, and never traded.

International prices

Every money figure is generated in US dollars and re-priced in your browser from your country's retail fuel and electricity prices and a parts-and-labour index, with the currency and units following the same setting. Change the country in the bar at the top of any page. The reference table is published at /assets/geo-prices.json.

Corrections

If a number here is wrong, it is wrong in the data or in the formula, and both are public. Write to corrections@motorjury.com with the page URL and what you believe the correct figure is; corrections are applied at the source so every affected page is fixed on the next nightly build.

What we never do

No fabricated numbers, no paid placement in verdicts, no AI-written filler prose. Advertising and affiliate links never touch a score. If data is missing we say so.

Questions: see about.