Read China's used-car market, trim sheet included

The seventeen-character VIN is not published. The record carries vin as null with vin_published false, and hands through the two opaque identifiers the source does use — a 32-character vehicle token and a 48-hex lookup signature — under names that say what they are. Neither is a VIN and neither decodes to one. If you join to a history or valuation database by VIN, this market cannot be joined that way: brand, series, trim id, registration month and mileage are the key you have. Our Korean engine does return the full VIN; this one cannot, and no amount of effort on our side changes that.

China quotes used cars in 万元 — ten thousand yuan. A car listed at 151.8 costs 1,518,000 yuan. Every row carries price_wan_cny with the source's own figure and price_cny with the multiplication done, and the same pair exists for distance: mileage_wan_km beside mileage_km. We did not pick one for you. A bare 151.8 in a price column next to a dollar auction feed is a ten-thousand-fold error, and it is the easiest mistake to make with this source.

The seller-type filter for private owners returns zero cars nationwide — on a genuine result page with all its filter furniture present, not an error page. Dealer and certified-dealer both return full pages, and they return disjoint sets of dealers. So this is the source's answer rather than a broken filter: its public listings are dealer-consigned. The value is still offered, labelled with that measurement, so nobody reads the empty result as a fault.

We asked for two cities that do not exist. Both answered HTTP 200 with a full page of cars from all over the country, which would have handed you national data labelled as one city's. The engine catches it by comparing your slug against the market the source says it applied, and refuses the call. Separately, a real city whose filtered set is thin gets padded with cars from neighbouring markets — that one is reported as city_padding rather than refused, with every row's own city id beside it so you can slice it yourself.

A detail record does carry the roadworthiness-test expiry, the insurance expiry, the warranty expiry where one exists, the number of previous owners, and the source's own sentence about major accident, flood and fire damage. The separate third-party condition report is a different thing: its endpoint answers reliably, but of nine listings sampled across three different parts of the site, not one had a report published — it covers a narrow inspected-stock segment. The response says report_published false in that case rather than inventing a clean bill of health, and the demand and condition indices beside it do come back every time.

We searched a full half-megabyte result page for a count. Every number-of-cars string on it belongs to a script template, a sidebar shop's own inventory, or a tag tooltip. There is no result total anywhere. total_reported is therefore null with a note saying why, and pagination.pages_available carries the one real signal the source does publish: the highest page its own pager offers.

Pages 1, 2, 3, 20 and 99 of one city shared zero listing ids with one another, 56 rows each. Page 100, page 101 and page 150 returned the identical ids. So one filter reaches about 5,600 cars and then stops silently. has_more is false at the ceiling, and the way deeper is a narrower filter, not a higher page number.

Cheapest-first returned 0.60, 0.68, 0.69, 0.72. Dearest-first returned 2050, 1288, 968. Lowest-mileage returned four cars at zero before the first at 0.001. Oldest-car returned 1991/01, 1991/04, 1993/03. Newest-listing returned a page published that same day. One further code the source offers returned 56 cars whose registration date was uniformly its unknown-date placeholder, so we cannot say what it orders and it is not offered. The consequence: there is no newest-car-first sort. Combine min_year with newest.

Every search row carries a factory trim id, and that id returns 71 named parameters from Autohome's model database — trim name, manufacturer, the price when new, body structure, class, engine, gearbox, fuel, seats and about fifteen grouped sections. The car's own equipment list is a separate 2.4 KB call. A trim does not change, so the spec sheet is cached for a week and costs a single credit.

An invented id answers NOT_FOUND with the source's own wording — that the car has been taken down — not an empty list and not a block. The message also says plainly that from outside, sold and never-existed cannot be told apart, so a sync can close the row instead of retrying it forever.

Dealer company name, numeric id, rating out of five and years as a member. No contact names and no phone numbers: the listing page does carry a named contact person and the parser drops it rather than publish it, which is also why there is nothing left to redact in the payload.

Titles, city names, colours, engine descriptions and every spec parameter come back in Chinese. Nothing is machine-translated, because a translation we invented is a value we cannot stand behind. Ids, prices, mileages and dates are language-neutral, and the city and brand directories give you both the Latin slug and the Chinese name so you can label your own interface.