How we measure
A valuation is only as good as its track record, so here is ours. On sales it has never seen, our estimate landed 40% closer to the actual sale price than the median for the same model at salvage auction, and 63% closer at retail, and our ranges contain the sale price 9 times in 10. Those are measured results, not targets: we also test each model on sales made after its training data ends, and a model that fails the rules we set in advance doesn't ship.
This page explains how the model reaches a number, what its range means, where our data comes from, how we test it, and where it is weakest. The full record is at the bottom.
How the model works
We value a vehicle in two markets: what it brings damaged at salvage auction, and what it would sell for undamaged at retail. Either way, the model finds the recorded sales most like your vehicle, all sold before the date being priced, and works out how damage, title, mileage and condition move the price. For a salvage value, it starts from what the vehicle was worth before it was damaged. It works the way an appraiser does: start from the closest comparable sales, then adjust for how your vehicle differs.
At its core this is regression analysis, with two newer techniques built in. The first is retrieval, the same idea modern AI systems use to look up evidence before answering: the model reasons from real, similar sales instead of relying only on what it learned in training. The second is conformal prediction, a branch of statistics for measuring uncertainty, which sets our ranges by testing them on sales held out of training. (Technically: retrieval-augmented quantile regression, with ranges set by split-conformal calibration on held-out sales.)
The result is a spread of likely prices, not a single number. The estimate and the range both come from that spread.
Find…
the past sales most like the vehicle you want to value.
Compare…
your vehicle with those sales to produce a spread of likely prices.
Test…
the range on sales held out of training.
An answer at every level of detail
You don't need every fact about a vehicle to get an answer. The model was trained to price vehicles from partial information, from a year, make and model alone up to a full description with a VIN and condition. Each detail you add changes what the estimate rests on. Usually the range narrows. Sometimes it widens, when a detail shows the vehicle is harder to price than it first looked. Either way, the range reflects what you've actually told us.
The comparable sales behind each estimate
Every report lists the past sales the model used, each with its date and price. These are the same sales the estimate rests on, not a separate list picked afterwards. The model doesn't simply average them. It weighs each one and adjusts for how that vehicle differs from yours, which is why an estimate can sit above or below most of its comparables. You can see exactly what the number is built on and judge it for yourself.
| Sale | Damage | Odometer | Sold | Price | Why it's comparable |
|---|---|---|---|---|---|
| 2020 Toyota Corolla LE | Front end | 86,573 | 2026-10-13 | $4,150 | same make & modelsame damage typeodometer within 4K misold 10 days ago |
| 2020 Toyota Corolla LE | Front end | 99,410 | 2026-10-13 | $4,250 | same make & modelsame damage typeodometer within 9K misold 10 days ago |
| 2020 Toyota Corolla LE | Front end | 101,336 | 2026-10-14 | $3,000 | same make & modelsame damage typeodometer within 11K misold 9 days ago |
| 2020 Toyota Corolla LE | Front end | 79,967 | 2026-10-13 | $3,750 | same make & modelsame damage typeodometer within 11K misold 10 days ago |
What the range means
Every estimate comes with a range that 9 out of 10 sales land inside, and a narrower range that about 5 out of 10 land inside. We don't assume those rates. We set each range on sales held out of training, then check it again on sales made after the model was built (for salvage, sales from mid-October; for retail, starting in November), separately for every level of detail and every price tier.
Where the data comes from
Our estimates rest on recorded sales, not asking prices. For damaged vehicles, that means final prices at the major salvage auctions: millions of sales from 2025 to today. For retail values, it means completed sales at large used-car retailers. We deliberately leave out the 2020–21 pandemic price spike, because prices from those years don't describe today's market. New sales arrive every day, and we retrain regularly, because a pricing model that stops learning goes stale within weeks.
How we test
We score each model on sales made after its training data ends, predicting each one as we would have before it sold. The comparable sales it may use all happened before that vehicle's own sale, so nothing from the future leaks in. We score every qualifying sale in the test period rather than a hand-picked set, and we compare against the simplest alternative: the median price of the same year, make and model. We also write down the pass-or-fail rules before we look at the results, and we don't change them afterwards. If a model fails them, it doesn't ship. Retail values get their first test of this kind in early November.
Where we are weakest
Some vehicles are harder to price than others. Our ranges are widest, and our estimates least certain, for very cheap vehicles, for rare vehicles with few recorded sales, and, at retail, for cars under $15,000. A wider range on these is the model telling you what it doesn't know.
The record
Claims are cheap, so here is the check. Each time we release a model, we score it on sales held out of its training and publish the results here. As sales made after its training data ends settle, we check it on those too. "Typical miss" is the median gap, in dollars, between our estimate and the actual sale price.
Salvage
With VIN
Without VIN
Share of sales inside the 9-in-10 range, by estimated value
| With VIN | Without VIN | |
|---|---|---|
| Under $1,500 | 89.9% | 89.9% |
| $1,500–$3,000 | 92.4% | 92.5% |
| $3,000–$6,000 | 91.4% | 91.7% |
| $6,000–$12,000 | 90.3% | 90.0% |
| Over $12,000 | 90.3% | 90.3% |
By level of detail
| Inside the 9-in-10 range | Inside the 5-in-10 range | |
|---|---|---|
| Year, make and model | 91.6% | 51.3% |
| + damage | 91.1% | 51.0% |
| + mileage | 91.4% | 50.9% |
| + VIN | 91.3% | 50.5% |
| VIN only | 91.5% | 50.7% |
| VIN + damage | 90.9% | 50.5% |
The range holds 9 times in 10 at every level of detail, from a year, make and model alone to a full description.
Measured on 60,000 salvage sales held out of training, sold January 1, 2025 to October 2, 2026. Typical miss is measured on 59,648 of them. On the most recent 4,536 sales, made September 15 to October 2, 2026, the typical miss was 40% smaller than a same-model median: $485 vs $806.
Retail
VIN, mileage and trim
VIN and mileage
VIN only
Year, make and model, mileage and trim
Year, make and model, and mileage
Year, make and model
Share of sales inside the 9-in-10 range, by estimated value
| VIN, mileage and trim | |
|---|---|
| Under $15,000 | 83.4% |
| $15,000–$25,000 | 89.7% |
| $25,000–$40,000 | 91.0% |
| Over $40,000 | 88.5% |
Measured on retail sales held out of training, sold July 9 to October 6, 2026: between 5,994 and 7,443 sales, depending on the level of detail. These sales fall within the training period. Our first test on retail sales made after training ends reports in early November, and its results will replace these.
Updated with each model release.
What we don't do
We don't hand-pick the examples we show, and we don't quote accuracy measured on sales a model was trained on. We don't republish raw auction listings: what you see are our estimates and the sales they rest on. And we don't hide the cases where we're less sure, because a range you can trust is worth more than a number that only looks precise.