CGEP: Pairing Chinese Generative-Search Citation Chains with an Open Paid-Placement Market

Generative engines answer product and service questions by citing web pages, which turns the citation chain into something worth buying. We release CGEP, a two-part resource for studying that market in Chinese AI search. CGEP-ITW pairs 124 generative-search answers over 62 stratified entities (1,777 citations, 768 cited pages, 684 landing-page snapshots) with a supply-side price list of 41,443 paid-placement offers across 11,355 domains, obtained unsolicited and released only as derived statistics. CGEP-Synth is a released synthetic benchmark of five manipulation techniques, each encoding coordination in a distinct substrate, with code and a datasheet. Our headline result concerns the field's default weak-supervision proxy. On real citation chains, low-authority concentration does not discriminate commercially motivated queries from motive-matched controls (odds ratio 0.67, p = 0.53): it fires on cheap commodities, whose citations are naturally obscure, and misses major brands, which buy placement on popular hosts instead. Supply-side purchasability does discriminate (OR 9.00, p = 8.4 × 10⁻⁶), and survives deleting every page on the twelve high-traffic platforms that dominate the citation surface (OR 8.04, p = 3.9 × 10⁻⁶). We also audit CGEP-Synth against itself and report exactly how much of its separability is guaranteed by construction, so that users of the resource can price its evidence correctly. All code, coded signals, derived crosswalk statistics, and the synthetic benchmark are public; the raw price list is withheld as dual-use. The dataset is released separately: CGEP v1.0, doi:10.5281/zenodo.22084160.