Teardown: What Actually Happens When You Check a Car's Price
Every day something like this lands in the crawl:
2020 Hyundai Tucson SEL AWD, 41,200 mi, clean title, one owner, $22,995. Listed in Columbus, Ohio.
Nothing special about it. That's the point. Most of the 1,108,535 listings we're carrying right now look exactly this ordinary — a make everyone knows (Hyundai's the fourth-biggest brand in our data, 60,937 listings), a body style everyone buys, a price that looks plausible at a glance. The question a buyer actually has isn't "is $22,995 a nice round number" — it's "is this a good price for THIS car." Answering that honestly takes more steps than people assume. Here's every one of them, in order, for this one listing.
The raw text
The listing arrives as a scrape, not a database row. "SEL AWD" is a trim string a dealer typed into a form, not a canonical fact. Some sources hand us mileage in km, some in miles, and a few omit it entirely and expect the buyer to call. About 41,200 of a Tucson's odometer digits are the single most load-bearing number in the whole pipeline, and it's also the one field most likely to be wrong, missing, or rounded to a suspiciously clean number. Before anything gets priced, it gets parsed — currency, unit, trim token — into a shape a model can actually use.
Make and model, unified
"Hyundai Tucson," "HYUNDAI-TUCSON-SEL," "Tucson 2.5L AWD" — three sources, same car, three different strings. If we priced each string independently, the model would think it had less data than it really does, and the price would wobble depending on which dealer's copywriter typed the listing. Every raw string gets mapped to a canonical family first. This is the least glamorous step in the entire system and also the one most likely to quietly wreck a price estimate if it's wrong — a mislabeled trim can shift a car into the wrong comparison set entirely.
Finding its cohort
A price only means something next to other prices. Our Tucson doesn't get compared against every Tucson on earth — it gets compared against Tucsons of the same model year, trim tier, and drivetrain, in the same region, because a used Tucson in Ohio and a used Tucson in Cairo are not competing for the same buyer and don't depreciate on the same curve. Region size matters here too: the more listings in a region, the tighter the cohort, the more confident the estimate.
USA sits mid-pack among our biggest regions at 48,127 listings — plenty of depth for a Tucson estimate, but nowhere near France's 67,569 or Japan's 61,940. A model priced off a thin cohort should carry a wider margin of error than one priced off a deep one, and ours does.
Mileage does the heavy lifting
Two Tucsons, same trim, same year, same region, $6,000 apart — nine times out of ten the answer is mileage, not haggling. It's the single feature the price model leans on hardest, which is exactly why sources that scrub or omit it cause the most damage downstream. When mileage is missing across an entire region's worth of listings, we don't guess blindly — we impute from the cohort's blended mileage curve and mark the estimate with an "≈" so nobody mistakes a modeled guess for a fact. Our Tucson has a real number, so it skips that flag entirely.
Trim and features, normalized
"SEL AWD" isn't just a badge — it's a bundle of options (leather-adjacent seats, blind-spot monitoring, larger wheels) that correlates with a real price premium over the base SE. The feature normalizer maps trim strings and equipment mentions into a small set of enums the model actually understands: transmission, drivetrain, fuel type, body style, and a trim tier pulled from a brand-sourced canonical map. Skip this step and the model can't tell a stripped rental-fleet Tucson from a loaded one — it just sees "Tucson," which is a much worse guess than it needs to be.
The number comes out
Everything upstream — canonical model, region cohort, mileage, trim tier, drivetrain — feeds into the trained price model, and what comes out isn't "the" price, it's a defensible range with a point estimate in the middle. For a well-populated region like the USA on a common trim like this one, that range is tight. For a thin region or a rare trim, it widens, and honestly, it should — a model that's equally confident everywhere is lying to you about one of those cases.
The deal check
Only after the estimate exists does the last step run: is $22,995 meaningfully below what the model expects for this cohort? Across the platform, 12,823 listings are carrying that flag right now — a below-market badge is a comparison against a mileage-and-trim-matched peer group, not against some Platonic "fair price" for a Tucson in general. If our example listing came in a few thousand under its cohort's estimate, it'd get flagged the same way. If it landed dead in the range, it wouldn't, and that's not a bug — most listings shouldn't be flagged. Deals are supposed to be the minority.
What's still a guess
Nothing in this pipeline sees the car in person. A rebuilt title hiding behind a clean-sounding description, curb rash the photos don't show, a timing belt nobody mentioned — none of that reaches the model unless it's in the text. The estimate tells you what this car should cost given what a Tucson like it usually is. It doesn't replace a mechanic, a Carfax pull, or looking under the mats. It just means you walk into that conversation already knowing what "usually" looks like, instead of guessing.
