Skip to content

Methodology

Data last built Oct 27, 2026. Source files are published daily by NHTSA.

Sources

Complaints, recalls and investigations are from NHTSA’s Office of Defects Investigation (ODI), public domain: consumer complaints (vehicle complaints only), recall campaigns (vehicle recalls only), manufacturer communications / technical service bulletins, and investigations. Crash-test ratings are from NHTSA’s New Car Assessment Program (5-Star Safety Ratings, nhtsa.gov/ratings). Fuel economy and annual fuel cost are from the EPA and Department of Energy (fueleconomy.gov). Files are matched to a make, model and model year by name; records that cannot be matched are left out. Each model-year page links to NHTSA’s own page for that vehicle (or to NHTSA’s recall search when the model name cannot be linked directly).

Counting complaints

Mileage

About half of complaints state a mileage. Mileage figures are those reported by the filer when the problem occurred. Medians and the histogram use only complaints with a mileage; mileages that are zero or implausibly high are ignored.

Parts

NHTSA’s component field is grouped into about 25 plain-English parts (for example “Engine”, “Brakes”, “Air bags”). Complaints that match none of them are listed under “Other or unknown”.

Complaint score (0–100) and comparison label

The score compares model years of the same model. Older model years were sold when fewer people filed complaints online, and newer ones have had less time to collect complaints, so raw counts are not compared directly. Instead, each year is measured against all vehicles of the same model year in NHTSA’s data, which allows for both the era and the time on the road. For each model year:

  1. Early share: complaints filed by the end of the second calendar year after the model year (roughly its first three years), plus 0.5 so that a year with none still has a small share, divided by the same count for all vehicles of that model year.
  2. Overall share: all complaints, divided by all complaints for vehicles of that model year. This catches problems that appear later, such as transmissions failing at higher mileage.
  3. Take the geometric mean of the two shares.
  4. Multiply by 1 + 2 × the share of complaints that report a crash, fire, injury or death.
  5. Multiply by 1 + 0.3 × the number of recall campaigns (counting at most 6) + 0.5 × the number of do-not-drive recalls (counting at most 2). A do-not-drive recall is also a recall campaign, so it counts in both. Recalls often cover several model years, which is why the count is capped.
  6. Multiply by a mileage factor, so that problems appearing early weigh more: 1.3 if the median mileage is 20,000 miles or less, falling evenly to 1.0 at 40,000; 1.0 up to 70,000, falling evenly to 0.8 at 90,000 and above.
  7. Divide by the median of the model’s compared years. This ratio is shown as, for example, “1.4× a typical year”. The score is 50 + 30 × log2(ratio), kept between 0 and 100, so 50 is a typical year for that model; half the typical rate scores 20 and twice the typical rate scores 80. Scores reach 0 at about 0.3× and 100 at about 3.2×.

Labels follow the score: 35 or lower (about 0.7× typical or less) is “Fewer complaints than typical for this model”, 65 or higher (about 1.4× typical or more) is “More complaints than typical for this model”, and anything in between is “Typical for this model”. The shading of the model year strip uses the same score. Only model years from 2008 that have at least 25 complaints are compared, and a model needs at least three of them. Model years are compared once their first three years of complaints are in; newer years are shown as “Too new to compare”.

An open NHTSA investigation is shown next to the label but does not change it: an investigation is a review, not a finding. The score is not adjusted for how many vehicles were sold, because NHTSA does not publish sales; a very popular model year will collect more complaints. Because each year is measured against all vehicles of its model year, an industry-wide wave of complaints (such as the airbag inflator recalls of the 2000s and 2010s) raises the baseline for every model of those years. For a model with only three compared years, the middle year is the typical year by definition. It describes complaint data, not how reliable a car is.

The box at the top of each model year page lists the figures behind the label: the ratio, how many complaints were filed in the first years, the most reported problem, recalls, and nearby model years of the same model with a lower score and no open investigation (not shown when the year itself already has fewer complaints than typical and no open investigation).

Trend

We compare the complaints filed in the last 3 full calendar years with those of the 3 years before. A ratio of 1.25 or more is “rising”, 0.75 or less “falling”, otherwise “stable”. Model years less than 4 years old are “too new”, and years with fewer than 15 complaints have “not enough data”.

“What owners report” summaries

For model years with enough complaints, an AI model reads a sample of up to 60 complaints and groups them into the problems owners describe. The sample is drawn at random, but complaints that state a mileage and complaints filed since 2020 are taken first, so it leans towards recent reports. Each problem must cite at least four of the sampled complaints it is based on; an automated check rejects summaries that cite complaints, recalls or service bulletins that are not in the source data, use verdict words such as “dangerous” or “defective”, state percentages, or contain personal information. The summary is labelled as AI-written on the page.

Counts per problem

To count a problem across all complaints, not just the sample, each problem gets a fixed keyword rule (for example: a transmission word and a symptom word such as “slip” or “shudder”). The rule is written with AI and tested on the sample: it must find at least 80% of the complaints the summary cites and no more than twice as many other sampled complaints. It is then applied to every complaint for that model year. A count is shown only if it is consistent with the sample (between 0.4 and 3 times what the sample suggests, and at least 5); otherwise the page shows the sample count instead. Mileage is the median of the matched complaints that state a mileage, shown with the median age of those same complaints (complaints filed in recent years often leave mileage blank, so this set leans towards earlier reports); when none states a mileage, the age shown is the median of all matched complaints; “owner-quoted repair” is the median of dollar amounts owners wrote in those complaints (shown from 5 quotes), as stated and not verified.

Example complaints

Example complaints are excerpts of about 150–600 characters. A complaint is left out if it contains something that looks like a vehicle identification number, a phone number, an email address, a web link, a city with state and ZIP code, a title such as “Mr” or “Dr”, wording such as “my name is” or “named”, or text NHTSA has already redacted. These checks are automatic and can miss things. The text is otherwise as filed, with these changes: text that is mostly in capitals is converted to sentence case (with the make and model names restored), quotation marks are dropped, character-encoding errors are repaired, NHTSA staff annotations (such as “*TR”) are removed, and text that was shortened ends with “…”. Each excerpt shows its public NHTSA complaint number. Under each AI-summarised problem, excerpts are taken only from complaints that the summary cites and that match the problem’s keyword rule, where one exists.

What this is not

Complaints are unverified, can be duplicated, and reflect what people chose to report. Nothing here shows what caused a problem or who is responsible, and none of it is legal or repair advice. See the disclaimer.