2026-08-28
A 6-month Google Maps switchback in 10 U.S. cities penalized congested roads and changed under 2% of trips. Target-road speeds rose about 2%; affected trip times fell 0.69%.
Turn-by-turn apps pick the fastest path for one driver. In a peak hour that path is often the same freeway, so the greedy choice piles traffic onto the bottleneck. Air traffic control and backbone networks already price journeys against the whole system. Road navigation has mostly stayed individualistic, even after fuel and tolls entered the cost function.
The theory is old. The Price of Anarchy bounds how much selfish routing can waste relative to a system optimum; Stackelberg routing and dynamic traffic assignment have the same message in richer settings. What was missing is a controlled test on a live city network, with real navigation users, not a simulated Wardrop equilibrium.
The intervention sits in the costing and ranking stage, after a set of feasible routes has already been generated. If a targeted segment is slower than a fixed fraction of its free-flow speed, every candidate that uses it picks up an extra externality cost ΔT. A cap cmax then limits how large that extra cost can grow, so a diverted route is allowed to be only a few percent slower than the best alternative. For trips that actually switched, the projected hit is about 30 seconds.
Segment lists are built two ways: peak density in the top quartile, or bottleneck status (the segment is slower than what sits immediately downstream). Road class is split on purpose. Atlanta, Boston, Los Angeles, New York and Philadelphia get freeways; Miami and Salt Lake City get arterials; Chicago gets a mix; San Francisco is pinned to US-101 bottlenecks; Seattle to I-5. Each city contributes roughly 80 to 120 segments. Treatment is a city-wide daily switchback, the same design ride-hailing firms use when they cannot split a marketplace in space. The headline estimator is hierarchical Bayes, with partial pooling across cities and hours. A mixed linear model plus 1,000 placebo reshuffles of the treatment calendar sits in the supplement as a check.
Drivers never see a "you were diverted" flag. Under 2% of observed trips get a different recommendation. Among trips that already touch a targeted segment, that share is about 5% to 15%. The feasible-path generator is untouched.
On the segments that received the penalty, the ten-city posterior median for speed is +1.77% for the full weekday window (+0.73 km/h) and +2.13% in peak hours. Los Angeles peaks at +4.56%; Atlanta at +3.30%. Fuel use, estimated with NREL's FASTSim physics engine plus a RouteE-style machine learning layer, falls 0.82% across cities for the full day.
"Affected" segments include both the roads traffic left and the roads it landed on, about 80% of each city's flow. Speeds there rise 0.37% over the full day and 0.49% at peak, with posterior probability 99.8% that the sign is positive. Trip-level travel times, available only for the last two months because of retention limits, drop 0.69% in the median for journeys that touch any affected segment, and about 0.5% to 1.0% city by city.
The geometry is unsurprising. Penalties sit on a few busy trunks; origins and destinations fan out, so substitutes spill onto a wider, thinner set of nearby roads. In Atlanta that looks like traffic peeling off the central I-85 corridor onto the I-285 loop. Histograms show a handful of segments losing a lot of volume and many others gaining a little.
The mixed model agrees: targeted-segment speeds are about 2.1% higher, and that point sits in the tail of 1,000 placebo calendars. Compliance barely moves. Completion and off-route rates stay within a percentage point; the 95% credible interval on a negative compliance effect is about ±0.006% in freeway cities.
| Metric | Setting | Median effect |
| Targeted-segment speed | 10 cities, all day / peak | +1.77% / +2.13% |
| Targeted-segment speed | LA / Atlanta peak | +4.56% / +3.30% |
| Affected-segment speed | 10 cities, all day / peak | +0.37% / +0.49% |
| End-to-end trip time | trips through affected roads | −0.69% |
| Share of trips rerouted | all observed trips | <2% |
Yearly CO2e, extrapolated from affected-segment fuel rates with a fixed penetration assumption, has a median cut above 1,000 tons in most cities (Los Angeles about 4,511; Chicago about 4,297; Seattle about 504). Every city's 95th percentile flips positive. City-level climate savings are a wide posterior, not a tight number.
This is a field measurement of system-aware routing on a production maps stack, with a switchback that can actually identify an effect. For anyone who ranks routes on a shared network, the recipe is concrete: add a capped externality in the scorer, move a thin slice of traffic, and the network metric is large enough to estimate. Thirty seconds and 2% are small locally. Scaled by city flow they become thousand-ton annual emissions in the model.
It is a proof that a marginal intervention works, not a claim that the penalties are optimal. The paper does not quantify how much more a fully dynamic penalty schedule could buy. Individual drivers will not notice: the saving is about 0.25% of an average affected trip, and about 1/40 of day-to-day trip-time noise. That is good for short-run acceptance and bad for hoping users will lock in a new equilibrium on their own.
Penalties track historical congestion, not a per-timestep dynamic assignment. Induced demand over a longer horizon, or a wider treated region, could eat the travel-time gain. Densities come from the maps pipeline, so roads that rarely need guidance are probably undercounted. Trip-level outcomes cover two months. Emissions are modeled, not measured at the tailpipe. Raw trip data stay confidential; only the analysis code is public. Every author is or was at Google Research.
Seattle's affected-segment speed effect is the softest of the ten, with only 70.8% posterior mass above zero for the full day. Arterial cities such as Miami and Salt Lake City have wider intervals on the targeted roads. Dropping holiday and severe-weather outliers makes the estimate more conservative: without that filter, peak affected-segment speed goes from 0.49% to 0.68%. They published the conservative cut.