How we score
Transparency is the whole product. This page is the entire formula — the same code runs for every car, and no human (and no advertiser) can nudge a result.
The data
Everything comes from the U.S. National Highway Traffic Safety Administration’s public databases: owner complaints (every defect report filed with NHTSA since the 90s), recall campaigns, and VIN decoding. Government data is public domain; we fetch it fresh, snapshot the raw response for reproducibility, and re-check every scored car at least every two weeks.
The Reliability Score (0–100)
Four sub-scores, each 0–100, combined with fixed weights:
| Sub-score | Weight | What it measures |
|---|---|---|
| Complaint volume | 40% | Complaints per year of the car’s life, on a logarithmic scale, divided by a coarse sales tier (see limitations) so a car that sold 400k/year isn’t punished for having more owners to complain. |
| Severity | 30% | Complaints involving a crash (weight 2), fire (3), injury (3), or death (6) count far more than squeaks and rattles. Any reported deaths subtract further. |
| Recall load | 20% | Recalls weighted by what they touch: airbags, fuel, brakes, steering, and seat belts count double; a “park it / park outside” order counts triple and caps the total score below Buy territory. The slope is gentle on purpose — a recall also means a free, documented fix exists. |
| Trend | 10% | Complaint rate over the last 24 months vs. the car’s lifetime rate. Early teething that faded scores high; problems still accelerating years in score low. |
The verdict
Buy at 68 or above · Caution from 42 to 67 · Avoid below 42. Thresholds were calibrated against cars with well-established reputations (good and bad) and are the same for every car.
“What breaks, and when”
We cluster complaints by NHTSA’s component category (engine, transmission, steering…). For mileage, NHTSA’s API doesn’t expose the odometer field — so we extract odometer mentions (“failed at 92,000 miles”) from complaint text and show the median only when at least three owners mention a mileage. It’s a rough guide, clearly labeled as such.
Owner ratings are a separate axis
Community ratings never touch the score above. If owners love a car the data says to avoid (or vice versa), we show you both signals side by side — divergence is information, not an error to smooth over.
Honest limitations
- No sales data. NHTSA doesn’t publish sales volumes and licensed sales data isn’t free, so we adjust complaint rates with a published three-tier table of US sales volume (best-sellers ÷ 3.5, high-volume ÷ 2, everything else ÷ 1). Coarse, but far better than pretending a Camry and a Lotus see the same traffic.
- Complaints are self-reported. They skew toward angry owners and toward model years that got media attention. The severity weighting and trend signal partially correct for pile-ons.
- Model-level, not car-level. A verdict describes the model-year population, not the specific car in front of you. Maintenance history dominates everything — get an inspection.
- Young cars are marked “early days”. A one-year-old model with few complaints isn’t proven reliable; we say so instead of handing out easy 95s.
Monetization, disclosed
The site may earn commissions from clearly-labeled partner links (VIN history, parts, service contracts). The verdict is computed before any link is rendered and no partner can buy a better score. That’s not a slogan; it’s the execution order of the code.