Methodology

This page documents how every number on this site is produced, including limitations and known biases. Everything is computed by an open-source pipeline from public data, and re-run every week. If you find an error, please tell us; corrections are listed at the bottom of this page.

On this page: Data sources · Driverless vs. supervised · How complete is the record? · Rates and comparisons · Map locations · Crash details · The pipeline · Download the data · Updates & corrections

Data Sources

NHTSA Standing General Order crash reports

Since 2021, the National Highway Traffic Safety Administration has ordered the companies it names — including every major developer of automated driving systems — to report crashes involving their vehicles. These reports are the base of this site: every crash shown here starts as a federal report filed by Waymo.

What must be reported has changed. Under the order's current version (the Third Amended Standing General Order, in force since June 16, 2025), a self-driving company must report within five days any crash involving a death, a hospital-treated injury, an airbag deployment, a pedestrian or cyclist, or a vehicle towed away; and by the 15th of the next month any other crash in which property damage is expected to exceed $1,000, or the self-driving vehicle struck another vehicle or object. Before June 16, 2025, any crash with any property damage had to be reported. In practice the change mostly removed minor bumps in which a Waymo was hit and the damage was small (Federal Register, May 2026).

NHTSA publishes two files: reports filed before June 16, 2025, and reports filed since. Both are re-downloaded every week (NHTSA still revises the older one). They include all reporting companies; we keep rows where the reporting entity is "Waymo LLC". NHTSA redacts exact dates (month only), addresses and coordinates.

Waymo Safety Impact Data Hub

Every quarter, Waymo publishes data files on its Safety Impact Data Hub, each covering September 2020 up to a cutoff about three months earlier. We use three:

• CSV1 — miles. Cumulative rider-only (no one in the driver's seat) miles per service area. Because every past quarterly file is still online, we download all of them to build a mileage time series.
• CSV2 — crashes. Waymo's own list of crashes during rider-only driving, linked to the federal report IDs, with exact dates, street-level locations and crash-type classifications.
• CSV3 — comparisons. Waymo's comparison of its crash rates with human-driver benchmarks, with confidence intervals.

Waymo's data currently runs through --; the federal record runs through --.

Which record is the base?

The federal record. Every crash Waymo reports to NHTSA enters this site within a week of NHTSA publishing it; Waymo's quarterly files only add detail (exact date, location, crash type) to crashes that are already there. Until that detail arrives, a crash has a month-level date and an approximate map location, and is labeled as such. An earlier version of this site did the reverse and, at times, silently left out nearly six months of recent crashes.

For the periods both cover, the two sources agree almost one-for-one, which is itself a useful check: Waymo's rider-only list matches the federal driverless record for all but a handful of crashes.

Driverless vs. Supervised Operation

Waymo's federal reports cover two kinds of driving: driverless (no human in the driver's seat; remote assistance doesn't count as a driver) and supervised testing (a safety driver behind the wheel, as in cities where Waymo isn't yet allowed to drive empty). All headline figures on this site cover driverless operation only: it is what the public rides in, and Waymo's mileage — the denominator for every rate — covers only driverless driving. The -- supervised-testing crashes stay in the data and can be shown on the map with the "Include supervised-testing crashes" toggle.

We classify each crash from NHTSA's "Driver / Operator Type" field. Two corrections apply: an operator type we don't recognize counts as supervised (keeping it out of the headline count), and a crash whose operator field is blank but whose narrative says a test driver was present — and which Waymo's rider-only list does not include — counts as supervised (this applied to two crashes in October 2026).

How Complete Is the Record?

• The newest months are still filling in. Companies have until the 15th of the month after a crash to report most crashes, and NHTSA publishes with a further lag. NHTSA's own "report submission date" field shows that 99.5% of Waymo's reports arrive within one month of the crash month, and that each file only partly covers the latest submission month. We therefore treat the two most recent incident months as incomplete: the record is currently considered complete through --. Charts shade later months.
• The June 2025 rule change breaks trends. Monthly counts and per-mile rates before and after June 16, 2025 are not like-for-like. Across that date, Waymo's rate of all reported crashes per mile fell from about 10 to about 5.5 per million miles, across Waymo's cities at the same time, while its rate of injury crashes — which are reported under both rules — barely changed. More than half of the drop is in contacts with a change in speed under 1 mph, the lightest kind of bump; the rest is consistent with other minor crashes the new rules no longer require, though some of it could be real improvement. We show both rates side by side so this is visible.
• Reports are filed by Waymo. The narratives and fields (speed, movement, injuries) are Waymo's account, sometimes based on what it was told by police or others. Injury levels are what was "alleged", for anyone involved.
• Duplicates. A crash involving two Waymo vehicles can produce two reports; this affects two pairs of reports in the data.

Rates and Comparisons

Crashes per million miles, by city

For each city where Waymo publishes mileage (currently Phoenix, the San Francisco Bay Area, Los Angeles, Austin and Atlanta), we divide the driverless crashes up to the month the mileage covers (--) by the cumulative miles over the same period. Counting newer crashes against older mileage would overstate the rate; an earlier version of this site did that. These cumulative rates mix the periods before and after the June 2025 rule change, and cities that launched recently have most of their miles under the newer, stricter-to-trigger rules — so compare cities with care.

Crash rates by quarter

Miles driven in a quarter are the difference between two consecutive cumulative releases of Waymo's CSV1. Crashes in that quarter are those in Waymo's CSV2 (the same rider-only driving the miles cover). Injury crashes are those Waymo flags as "any injury reported"; police-reported crashes are flagged similarly. Quarters where a city has less than a million miles are left out.

Comparisons with human drivers

The "fewer crashes than human drivers" figures are Waymo's own, read automatically from its CSV3 file each quarter. We use the comparison Waymo itself headlines, against a "dynamic" human benchmark that weights human crash rates by where in each city Waymo drives. The method is described in a peer-reviewed paper by Waymo researchers (Kusano et al., 2025, Traffic Injury Prevention) and follows an industry checklist for such studies. We show the 95% confidence intervals Waymo publishes, and the underlying crash counts.

Limitations to keep in mind: the analysis is by Waymo, about Waymo; human benchmarks can't fully match Waymo's roads, hours, weather and vehicles; and the rarest outcomes rest on very few crashes (serious-injury-or-worse: --; fatal: --). One crash involving two Waymo vehicles, with a reported life-threatening injury to another driver, is counted by Waymo as "unknown" severity pending a police report; Waymo notes that counting it would raise its serious-injury rate. An independent study by the Insurance Institute for Highway Safety (July 2026) found smaller but still large reductions (68% fewer police-reportable crashes, 81% fewer injury crashes per mile, 2021–2024).

Map Locations

NHTSA redacts addresses and coordinates. Locations come from the street description in Waymo's CSV2 (for example "Florida Street near 24th Street"), which we geocode with OpenStreetMap's Nominatim service, searching in the crash's own city. Every marker carries a precision label, shown in its popup:

• Intersection or address (-- driverless crashes): the full description matched.
• On the right street (--): only one of two named streets matched, so the marker is on the correct street near, but not exactly at, the crash.
• ZIP code center (--): the address couldn't be placed inside the ZIP code Waymo reported, so the marker sits at that ZIP's center.
• City center (--): crashes Waymo's quarterly file doesn't cover yet, placed near the center of the city in the federal report with a small random offset.

ZIP check. Until June 2025, Waymo also published each crash's ZIP code. We use U.S. Census Bureau ZIP-area centers and sizes to check that a geocoded marker actually falls inside that ZIP; if it doesn't, we search again within the ZIP's area only. An October 2026 audit found that before this check, about a quarter of "on the right street" markers were placed well outside their reported ZIP code. Waymo no longer publishes ZIP codes for newer crashes, so those can't be checked this way.

Geocoding results are cached in the repository so markers don't move between updates.

Crash Details

• Speed — "SV Precrash Speed (MPH)", the Waymo's speed just before the crash. Bucketed into 0, 1–5, 6–15, 16–25, 26–35 and 36+ mph.
• Stopped or parked — "SV Pre-Crash Movement" is "Stopped" or "Parked".
• Hit from behind — for rear-end crashes, NHTSA's contact-area fields show whether the Waymo's rear (and not its front) was struck.
• Kind of road — "Roadway Type" (street, intersection, parking lot, highway, traffic circle). Until October 2026 this site guessed the location type from words in the narrative, which badly overcounted parking lots and intersections; see corrections.
• Crash type — Waymo's classification in CSV2 (e.g. "V2V F2R", translated as "rear-end collision"). Available only once Waymo's quarterly file covers a crash; pedestrian, cyclist and motorcycle counts fall back to NHTSA's "Crash With" field for newer crashes.
• Injury severity — NHTSA's "Highest Injury Severity Alleged" (none, minor, moderate, serious, fatal). The "any injury", "police-reported" and "airbag" flags come from Waymo's CSV2 and are computed among crashes it covers.
• Time of day — NHTSA's incident time. Day of week needs an exact date, so it covers only crashes in Waymo's quarterly files.

Fault. Federal reports don't assign fault, and neither do we. A Waymo that was stopped when hit, or a "rear-end" crash, describes what happened physically, not who was responsible. The short summaries of fatal crashes on the home page draw on police and news reports as well, and are written by an editor.

The Pipeline

An automated job runs every Monday on GitHub Actions. It runs the test suite, then six steps:

1. Download both NHTSA files and Waymo's CSV1, CSV2 and CSV3 (the newest quarter is detected automatically), plus every past CSV1 release.
2. Merge and clean: keep the latest version of each report, classify driverless vs. supervised, attach Waymo's details, assign each crash to a metro area (crashes in unrecognized places land in an "Other" bucket — never dropped).
3. Compute statistics: every number on the site.
4. Map data: geocoding, the map file and the downloadable CSV.
5. Serious incidents: moderate-or-worse injury crashes with their narratives.
6. Editorial checks: compare the data with what an editor has already reviewed. A new fatal or serious-injury crash, a crash in an unmapped place, or a new Waymo release opens a review issue on GitHub — so sentences an editor wrote can't silently go out of date.

The output is checked before publishing (crash counts must never shrink or go back in time). The code is open source; to reproduce everything, clone it and run make data, then make serve.

Download the Data

waymo_crashes.csv — every crash on the map (driverless and supervised), one row per crash, rebuilt every week. Columns:

The original sources: NHTSA Standing General Order data and the Waymo Safety Impact Data Hub downloads. The mileage milestones from Waymo announcements are in the repository (data/static/mileage_milestones.json), each with its source.

Updates & Corrections

Corrections

October 23, 2026 — A full review of the site's data, methods and text found and corrected these errors:

• The FAQ described the fatal crashes as "both" cases after a third, in Dallas in August 2026, had entered the data. All three are now summarized individually, and an automated check flags any new fatal crash for review.
• The "location type" chart, built from words in the crash narratives, said 45% of crashes were at intersections and 26% in parking lots. Using the structured field on the federal reports, the shares are about 22% and 9%. An FAQ answer about parking-lot crashes based on the old figure was removed.
• Per-city crash rates counted crashes up to three months newer than the mileage they were divided by, overstating every rate by roughly 5–10%. Atlanta's published mileage was missing.
• Waymo's "fewer crashes than human drivers" figures were presented as covering all of Waymo's miles and attributed to a 2025 paper; they came from an older data release. They are now read from Waymo's latest release, with confidence intervals.
• Two crashes during supervised testing (Washington, D.C., March 2026; San Francisco) had been counted as driverless because a field was blank on the federal report.
• Roughly 330 map markers had been placed well away from the crash; markers are now checked against the reported ZIP code.
• Several facts were out of date: Waymo serves 15 metro areas, not 11; Denver is open to the public, while Washington, D.C. and Philadelphia allow only supervised testing; Waymo's San Antonio service was paused for about five months in 2026, not two weeks; and a June 2026 recall was missing.
• Two mileage milestones were misdated; the reporting threshold described on several pages was imprecise; the latest months of federal data were presented as complete.

A detailed, technical log of the review is in the repository (OVERHAUL_LOG_2026-10.md).

Data updates

What each automatic weekly update changed:

Counts before October 23, 2026 include two supervised crashes later reclassified (see corrections).