Source rules
- Every fact links to its source and carries an as-of date.
- Open data, licensed data and opt-in rider data only. We never scrape Waymo, Uber, Lyft or Tesla apps or private APIs; their terms forbid it.
- Curated facts are kept in version-controlled files, so every change has a history.
- Where a value is modelled, we show a range and a confidence label.
Method and confidence labels
- Reported
- Published by a regulator, company or third-party study.
- Observed
- Computed by us from raw public data, such as trip records.
- Modelled
- Estimated by us where no direct figure exists.
- High
- Primary source: regulator, statute or the operator itself.
- Medium
- Reputable secondary source, or a primary source more than a year old.
- Low
- Small samples, sources that disagree, or our own inference. Treat as a lead.
Service status
Each operator in each market has one status, from least to most advanced. A market shows its most advanced status.
- Announced
- Testing
- Supervised
- Driverless, invite or free
- Driverless, paid
Regulatory openness rubric
Regulation is rated per jurisdiction: a US state, or a city or emirate elsewhere. Pending bills are tracked as an outlook flag, not folded into today's level.
| Level | Meaning |
|---|---|
| 5 Open | No discretionary permit, no fleet cap |
| 4 Permitted at scale | Permit with large fleet caps |
| 3 Gated commercial | Driverless paid service under case-by-case permits |
| 2 Legal path, not yet live | A permit route exists but no driverless paid service |
| 1 Supervised testing only | Safety driver required |
| 0 Prohibited | Banned |
Modifiers, capped at ±15 in total:
- +5 state bars cities from regulating avs
- −5 live labour or local-control threat
- −5 to −10 for a regulator freeze in the last 12 months
- −5 for the EU/EEA cap of 1,500 registrations per vehicle type a year
- ±3 for airport access
Potential score
The Potential score rates the 100 largest US metros 0–100 from public data (see the rankings). Each metric is ranked as a percentile across the 100 metros; each sub-score is the mean percentile of its metrics; the weighted sum is then multiplied by a legal gate: 1.0 where driverless paid service is allowed (levels 3–5), 0.7 at level 2, 0.4 at level 1 and 0 at level 0. Grades: A at 85 and above, B 70–85, C 55–70, D 40–55, F below 40.
A score is as confident as its least certain input. Where a state's regulation level is our low-confidence estimate, the market page also shows the score with the level one step lower and one step higher. Back-tests check that San Francisco, Phoenix and Los Angeles rank in the top 15 before the gate, and that New York, Boston and Washington rank in the top 20 before it and grade D or lower after it.
| Sub-score | Weight | What it measures |
|---|---|---|
| Demand | 30% | Households without a car, metro population, population-weighted density, residents aged 18–34, airport passengers per resident |
| Revenue | 20% | Median household income, driver hourly pay |
| Competition | 15% | Held at 50 for every metro: its measures (robotaxis per ride-hail trip, operators, wait times) are observed service data, which Live scores use |
| Operability | 15% | Days with snow, heavy rain and fog |
| Regulation | 20% | Openness rubric level plus modifiers (capped at ±15) |
Live score
Metros with 90 days or more of paid driverless service (counted from a sourced launch date) also get a Live score, ranked against each other. It keeps the fundamentals and replaces the modelled proxies with observed service data: the fare for a 5-mile, 15-minute trip from rate cards, airport passengers weighted by robotaxi airport access, the number of driverless operators, a low-price entrant flag, and robotaxi injury crashes per million residents over the last 12 months. Each metro gets two scores: one for lidar fleets (Waymo, Zoox, with the Potential weights) and one for camera-only fleets (Tesla), where operability counts double. Crashes grow with fleet activity, so long-running markets score lower on regulation; see the full methodology for known limits.
Owner-yield calculator
The calculator estimates what one robotaxi could earn its owner per year on a ride-hail network: 365 × trips per day × (fare × (1 − network cut − manager fee) − miles per trip × operating cost per mile). Tesla has published no revenue split, so every input is an assumption with a low, base and high case, each shown with its source and confidence. Results are estimates, not forecasts.
| Input | Low | Base | High |
|---|---|---|---|
| Trips per day | CPUC-based estimate, low end | CPUC-based estimate (Waymo California) | CPUC-based estimate, high end |
| Fare per trip | Tesla's Austin card before March 2026 | Tesla's rate card (the market's own, else Austin's, low confidence) | Waymo's base and per-mile rates in the market |
| Network cut | 25% (Musk, 2019) | 30% (Musk, 2019) | 30% |
| Manager fee | 0% | 0% (owner self-manages; assumption) | 20% (assumption; no published rates) |
| Miles per trip | Lowest month, CPUC | 12-month average, CPUC (paid + empty) | Highest month, CPUC |
| Operating cost per mile | $0.20 (Musk's long-run Cybercab target) | $0.74 (Morgan Stanley, Tesla, April 2026) | $1.36 (Morgan Stanley, Waymo) |
Fares are priced for the paid miles of an average trip. Trips per day and miles per trip are Waymo California figures applied everywhere; Tesla's fleet carries far fewer riders today. Morgan Stanley's cost figures exclude safety-monitor pay and up-front AV development. The low and high results combine every input's worse and better case; with a negative margin per trip, more trips make the result worse.
Rider panel
Operators do not publish what rides cost or how long riders wait, so riders can share their ride receipts on the rider panel. Each receipt is uploaded with explicit consent, read by an automated service (Anthropic's Claude) for the fare, distance, ride time, wait, operator and date, and then deleted; an image is never kept more than 7 days. Receipts that are not robotaxi rides or cannot be read are not used.
Figures are published per city, operator and month only once receipts from at least 30 different riders are in (each rider counts for at most three, and the same image shared twice counts once), never for a single ride: the median fare per trip (with the middle half of fares, 25th to 75th percentile, as its range), the median fare per mile and median wait when 30 or more riders' receipts show them, and the number of receipts. They are labelled observed, medium confidence: riders choose to share, so the sample may not match all rides. US dollar receipts only for now. See the privacy notice.