Applies to every study on CarsMultiverse. If you want to reproduce a number, this page tells you how.
Every study here is rebuilt by a scheduled job that reads a public register. Nothing is a summary of someone else’s summary. Each page states which file it read and when, and every study ships a CSV so you can check the arithmetic yourself.
What we read
| Source | What it gives us | Read |
|---|---|---|
NHTSA bulk recall file (FLAT_RCL_POST_2010) |
Every American recall campaign since 2010: campaign number, make, model, year, type, units, dates | hourly |
| NHTSA recall notice text | Summary, consequence and remedy wording for tracked campaigns | every 6 hours |
NHTSA recallsByVehicle |
Per-model campaign counts, used for cross-country comparison | hourly |
| NHTSA annual completion-rate report | Repair rates per campaign: published once a year, as a PDF, in no other form | daily |
| the Netherlands vehicle register (Netherlands) | Dutch recall notices and fleet counts | daily |
Live rebuild times and source health for all seven studies are on the status page, which is generated, not written.
How we prove wording changed
Each time we read a notice we take a SHA-256 hash over its remedy, summary and consequence, joined by newlines, and store it with the timestamp of the read. When a later read produces a different hash, the notice was edited, and we publish both versions with the hour each was seen.
The dated manifest is downloadable. Keep a copy and you can prove later that the wording changed without having to trust our word for the earlier version, which is the point of publishing it.
We can only prove a change we witnessed. The archive begins on 17 August 2026. For anything edited before that date we hold no earlier version and make no claim.
Counting rules
- One campaign counts once. Counts are of distinct campaign numbers per make and model. A single campaign covering six model years is one recall, not six.
- Vehicle recalls only. Equipment recalls are excluded throughout. Mixing the two inflates every figure, and equipment campaigns are not comparable between countries.
- Model-year window. The per-model index covers 2016–2025 on both sides of any cross-country comparison, so the two are counting the same span.
- Manufacturer names are merged, and the merge is published. Registers spell one company many ways: NHTSA carries both “Chrysler Group LLC” and “Chrysler (FCA US, LLC)” as separate entities. We flatten these to one maker and every CSV carries the raw spellings we merged, so you can disagree with a merge and redo it.
What we exclude, and why
- Dates that contradict their own campaign year. A campaign numbered for 2018 carrying a 2011 report date is a data-entry error, not an eight-year delay. Rows failing that consistency check are dropped from the delay study and the count of dropped rows is stated on the page.
- Models that are not comparable. Vans, campers and chassis conversions are excluded from cross-country comparison, because the naming does not correspond between registers.
- Vehicles absent from a register. If a model is not sold in a country, “zero recalls” there is meaningless. Those are published as not comparable rather than counted as zero.
- Fleets too small to mean anything. Cross-country rates are only computed above a minimum national fleet size.
Two things we learned the hard way
These are recorded because they affect numbers we published, and because anyone reproducing our work will hit them.
NHTSA answers “no recall for this vehicle” with an HTTP 400. The body of that 400 reads {"Count":0,"Message":"Results returned successfully","results":[]}. The status line and the payload disagree. We read the payload: an empty answer is an answer, not a failure. Treating the status line as authoritative caused this site to report the source as degraded on days it was replying normally, and that error is fixed.
The completion-rate PDF is not text. It is exported with subset fonts, so the page contains glyph identifiers rather than characters. We rebuild the character mapping from each page’s own font tables, track the coordinate transforms so figures land against the right campaign, and refuse to publish a year unless at least 95% of the rates align to a campaign baseline. If that check fails, the year is not published.
When a source fails
A study that cannot read its source does not invent a number and does not quietly serve an old one. It keeps the previous figures, marks itself on the status page, and states when it last succeeded. Current there means the study rebuilt within four times its promised interval: deliberately generous, because one missed run is not a scandal and we would rather under-claim a problem than cry wolf.
Corrections
Errors get a dated entry on the corrections page saying what was wrong and what changed. We do not edit a number quietly. If you can show a figure here is wrong, tell us: that is the most useful thing anyone can send us.
Reuse
Recall text belongs to the authorities that published it. The dated archive, the joins between registers and the studies built on them are ours, and the CSVs on the data page are free to reuse with attribution to CarsMultiverse.
Three counts of the American register
Asked how many American recall notices this site holds, there are three defensible answers, and they differ because they are answers to three different questions. This site quotes the first.
- 15,236 held. Every campaign whose fingerprint we hold, including ones the latest published file no longer lists. This is the figure on the front page and in the band of registers, and it is what we mean by held.
- 15,232 in the file. The campaigns present in the regulator’s own file at the last read. Lower than the first, because a campaign can leave the file after we have read it, and we do not throw away what we already hold.
The register was last read 10 September 2026. These three figures are counted when this page is rebuilt, not typed in.