Coordinates shift hundreds of meters on Chinese basemaps

• A point collected at 116.397428, 39.909230 in Beijing appears at 116.403672, 39.910634, 555.6 m to the north-east.
• The offset measures 265.3 m in Urumqi, 481.8 m in Shanghai and 621.7 m in Guangzhou, and its direction rotates across the country.
• Every POI in a batch moves together, so a single point looks like a data-entry error until the whole layer is checked.

1. Identify the source CRS: device and government open data are usually WGS84, Amap API output is usually GCJ-02.
2. Identify the basemap CRS: Tianditu is CGCS2000, Amap and Tencent are GCJ-02, Baidu is BD-09.
3. If the two match, do not convert; if they differ, convert exactly once before rendering.
4. Convert with gcoord.transform(data, 'WGS84', 'GCJ02') or an equivalent implementation and inspect one known point.
5. Never hand-tune individual points after conversion; that turns a systematic error into random noise.

Run a single datum conversion on the way in, e.g. gcoord.transform(geojson, 'WGS84', 'GCJ02') for Amap/Tencent tiles, or WGS84 to GCJ-02 to BD-09 for Baidu. Reverse GCJ-02 to WGS84 has no analytic solution, so iterate until the residual is under 1 m.

• Applying the popular 'add 0.0065' constant to GCJ-02: that constant only approximates the BD-09 step and leaves the 480 m GCJ-02 offset untouched.
• Converting twice because both server and client do it, which doubles the offset instead of cancelling it.