Two Chinese used-car markets, one in yuan and one in dollars

Guazi sells nationwide into each city, so a search scoped to Beijing legitimately returns cars sitting in Shenyang, and one scoped to Shanghai returned cars from nine different cities. The filter is not broken and the rows are not wrong — the city you pass is the buyer's market. Every row carries its own city, which is where that car actually is. If you need strictly one location, filter on the row's city, not on the request.

On the car we measured, the page's structured data says Shenzhen while the car is in Suzhou — and guazi's own plain-text mirror of the same page agrees with Suzhou. One field is the browsing market, the other is where the car is. We publish the car's real location as city and expose the other as market_city, so the difference is visible instead of quietly wrong in your database.

The seventeen-character number appears nowhere on a domestic listing — not under a Chinese label, not under an English one. So vin comes back null with a flag saying the source does not publish it, rather than filled with something that looks like a VIN. On the export side guazi publishes the first three and last four characters; the middle is not on the page at all, so that mask is the honest maximum and it arrives as vin_masked. The identifier you can actually join on domestically is the listing id, which round-trips unchanged between search and detail.

A fixed-price export row publishes an FOB asking price. An auction row publishes the CURRENT BID in the same field — we measured thirty dollars on a 2023 Hyundai Palisade with 39,900 km, while every other facet's cheapest car that minute was between three and thirty-four thousand. Reading that as a market price would be wrong by three orders of magnitude, so the two are split: price_usd for an asking price, current_bid_usd for a live bid, and price_kind naming which one you are holding.

A domestic detail record carries the number of insurance claims against the car and the number of times it has changed owner, guazi's condition grade, an appearance score out of a hundred and whether an official two-hundred-point inspection report exists. An export record goes further and states no-accident, no-flood and no-fire explicitly, each as its own flag, next to an A-to-D grade. These are the fields an importer actually buys, and they came back on every car we sampled.

China quotes used cars in 万 — units of ten thousand yuan — and distance in 万公里. A listing reading 14.43 and 10.94 is 144,300 yuan with 109,400 km on it. Nothing in this engine returns the source's bare figure: prices arrive as absolute yuan and mileage as absolute kilometres, and the price bands are converted too, so band price12 reports 130,000 to 180,000 rather than 13 to 18. The export side carries the dollar price alongside guazi's own yuan exchange rate, so you can reconcile the two markets yourself.

Forty rows a page, pages one to twenty, and page twenty-one returns nothing on every facet we tried — even where the source's own headline count was forty-three thousand. The way deeper is narrower: brand, then model series, then price band. A series-plus-band facet returned seven cars and said so, which is what a real narrow result looks like rather than a padded grid. The export side is tighter still: twenty rows per facet and no second page at all, because its own filter and page links do not work without running its JavaScript. We expose none of those parameters rather than accept one the source would silently ignore.

Brand returned only that marque; model series only that series; each price band's rows all fell inside the yuan range guazi declares for it, on four separate bands. The filters the source silently ignores — body type, fuel, year, an arbitrary price — are not offered at all, because a filter that is accepted and then dropped is worse than one that does not exist.

A city slug that does not exist returns NOT_FOUND. A real city with nothing in stock returns success with an empty list and a reported total of zero. Those are different facts and they are reported differently, so a typo cannot read as 'guazi has no cars there'. A dead listing id likewise returns NOT_FOUND with the source's own refusal behind it, which lets a sync close the row instead of retrying it.

Guazi publishes its own crawler index of every live domestic listing: thirteen pages of roughly ten thousand ids, each with the date it was last updated — about 127,000 cars. Pull the ids, diff the dates against what you stored, and spend a detail call only on what moved. It is priced by the rows it returns, so a full page costs five credits rather than the ten thousand a one-by-one walk would.

Domestic titles, cities, colours and engine descriptions come back in Chinese as guazi publishes them; nothing is machine-translated, because a translation we invented is a value we cannot stand behind. The brand and city directories give you both the Latin slug and the Chinese name. The export marketplace is written in English by guazi itself, including the options sheet, which is why it is the easier of the two to put in front of a non-Chinese-speaking team.