Purchasable, Not Obscure
Ask a Chinese AI search product whether a rice cooker is worth buying and it answers in prose, citing a handful of pages. Those citations are the answer's evidence. If a place on a cited page can be bought, part of the answer can be bought too. The practice has a name, generative engine optimization (GEO), and in China it has a visible market: placement on content platforms is sold per domain, and some offers advertise "GEO" or "AI inclusion" as a feature.
Work on detecting this kind of manipulation usually treats the manipulation itself as unobservable and relies on a proxy, most often the share of an answer's citations that land on low-authority domains. That proxy has not been tested against a control group that asks the same kind of question and differs only in commercial motive. My paper, CGEP: Pairing Chinese Generative-Search Citation Chains with an Open Paid-Placement Market, builds that test. It releases a two-part resource (dataset DOI): real generative-search answers joined to the supply side of the placement market (CGEP-ITW), and a synthetic benchmark that comes with an audit of its own construction (CGEP-Synth).
A control that asks the same question
The unit of observation is a citation group: all pages cited by one answer. I sampled 62 entities in seven strata and asked two questions about each, "is X worth buying" and "X or Y, which is better", for 124 answers collected on 11 July 2026. They contain 1,777 citations over 768 distinct pages on 334 hosts; 684 of those pages were fetched as snapshots.
Four strata carry commercial motive: small appliances and electronics (A), education and certification (B), medical aesthetics and health products (C), and major brands (D). The control that carries the design is E1: seven low-ticket household goods, a drying rack or a toilet brush, asked with the same two questions. Relative to the treatment strata the question type stays fixed and one thing changes: whether buying placement would pay. A factual-question control would change the question type as well as the motive. E2 (three factual entities) and Dcal (two non-Chinese brands) serve as floors.
The supply side is a paid-placement price list received unsolicited in July 2026: 41,443 offers across 11,355 domains, of which 816 offers on 204 domains are sold explicitly in terms of generative retrieval, at a median price of 18 RMB. Nothing was bought or negotiated. The raw list is an operational directory of placement channels, so it is withheld as dual-use; the release contains only derived statistics (domain-level overlap counts, price quantiles, GEO-flag counts), with overlapping domains pseudonymized.
Each group is coded on seven signals tied to a five-technique taxonomy (below). Two carry the result. S1 fires when at least half the citations sit on domains outside the Tranco top 100k: the low-authority proxy. S7 fires when a cited host appears in the price list, a table lookup with no coder judgment.
Low authority does not discriminate; purchasability does
≥ 50% of citations on domains outside the Tranco top 100k
Odds ratio 0.67, one-sided Fisher p = 0.86. Continuous share 0.325 vs 0.351, p = 0.53.
Start on S1. It fires on 25 of 100 treatment groups and 8 of 24 controls, an odds ratio of 0.67 in the wrong direction and not significant (p = 0.86; on the continuous share, 0.325 against 0.351, p = 0.53). Switch to "by stratum" to see why. E1 trips S1 on 8 of 14 groups: a drying rack is answered from small retailer and manufacturer pages whether or not anyone paid for them. Major brands (D) trip it on 2 of 20, because they buy placement on popular hosts. Popularity measures salience, and salience is only incidentally related to manipulation.
Now S7. Three quarters of treatment groups cite at least one purchasable host (75 of 100) against a quarter of controls (6 of 24), odds ratio 9.00, p = 8.4 × 10−6. Restricting to offers sold explicitly as GEO raises it to 11.67. The obvious objection is that Chinese AI search leans on a few large publishing platforms that appear in any placement catalogue; three such hosts account for 151 of the 245 cited pages that overlap the list. Deleting every page on the twelve highest-traffic platforms leaves 66 of 449 treatment pages (14.7%) against 3 of 143 control pages (2.1%), odds ratio 8.04, p = 3.9 × 10−6. The gap does not come from the portals.
Stratum D is the case a single detector misses. Its S1 rate is the lowest in the sample and its S7 rate is 12 of 20, every hit on an explicit GEO offer. Across the cross-tabulation, the most common cell in the commercial strata is S1 absent and S7 present: 23 of 30 groups in A, 17 of 20 in C, 12 of 20 in D. Content placed on popular hosts is invisible to a detector that equates manipulation with obscurity.
A note on method. An earlier version of the coder called any host missing from a hand-built whitelist low-authority. On the full sample it classed the China Meteorological Administration and a Chinese Academy of Sciences institute as marketing sites and fired on all 20 groups in E1 and E2. The current definition uses a public ranking (Tranco list N2Q7W, threshold fixed before analysis). The release ships the rank subset for the 288 cited domains rather than the list.
The taxonomy, and the technique that never appeared
The paper sorts manipulation on two axes: the channel through which the engine is influenced (retrieval corpus, runtime session, or authority signal) and the organizational form (single item or coordinated cluster). Five techniques result:
- T1, account-matrix farming: many accounts run by one party publish same-leaning content.
- T2, citation stuffing: pages cite one another to manufacture authority.
- T3, parasite content: commercial content on a popular, trusted host.
- T4, hidden injection: page content addressed to the model rather than the reader.
- T5, fake-review clusters: coordinated reviews across platforms.
The taxonomy is descriptive, drawn from public enforcement reporting and from the offer categories in the price list. The defender's task splits into detection (is this chain manipulated), classification (which technique), and attribution (which actor cluster placed it). CGEP-ITW supports only the first two, and only as observability, because no ground truth about transactions exists in the wild.
One cell stayed empty. No LLM-directed hidden instruction appeared in any of the 684 snapshots, so S5, the T4 signal, is zero throughout. On the synthetic benchmark the same technique is perfectly separable because a boolean flag carries it. The easiest technique on constructed data is the one I could not observe on real pages.
A synthetic benchmark that audits itself
CGEP-Synth generates campaigns over cluster sizes 3, 10 and 30 crossed with three levels of intra-cluster homogeneity, in three realizations up to all five techniques (7,740 items, 774 per technique, balanced against organic items), each under five seeds. Items carry accounts, sellers, platforms, scalar and framing features, and edges in an interaction graph and a citation graph. It contains no natural-language text.
Its organizing thesis is that each technique hides coordination in a different substrate, so a detector on the wrong substrate fails. That thesis is built into the generator, and the paper states how much of the resulting separability follows from construction:
- Of the twelve technique-by-substrate cells, eight are structural zeros: the objects a detector needs are absent from that substrate, so 0.00 reflects an empty graph, not a failed model. All 1,501 interaction edges lie inside T1 clusters.
- Two more are near-perfect because a label is readable as a feature. The DOM-carrier flag is true for exactly the 774 T4 items. Host authority is a constant 0.300 for four techniques, while T3 draws from 0.701 to 0.950 and organic items from 0.2 to 0.6.
- Only content similarity on T3 (F1 0.547) and T5 (0.351) is earned by modeling, and there the classes overlap.
What remains useful is a harness for attribution methods:
Content features tell manipulated items from organic ones but not one seller from another. The interaction graph holds the seller structure because the generator puts every interaction edge inside a T1 cluster.
Content features detect coordination well (F1 0.934) but attribute it to a seller cluster at 0.185, while Louvain communities on the interaction graph reach 0.959, both under one-to-one Hungarian assignment (a majority-mapping variant inflates every graph method to 1.00). A graph autoencoder does not beat the classical baseline. These numbers say a method can or cannot recover a structure the generator planted. Claims about real accounts have to come from somewhere else, which is the reason CGEP-ITW exists.
Evidence strength and boundaries
CGEP-Synth is synthetic and mostly decided by construction. Ten of its twelve technique-by-substrate cells are fixed by the generator. It supports comparing attribution methods, not statements about real sellers.
CGEP-ITW is observational and small. It is one instrument at one point in time, and generative-search citation behavior shifts over months. The instrument is a search-grounded LLM API with live retrieval over Chinese pages, not the consumer apps of Chinese engines, which need authenticated sessions; the bare chat APIs of those vendors do not browse and, in pilot testing, invented source URLs. The dataset shows how the Chinese content ecosystem appears in a generative-search citation chain, not how a particular product behaves. Coding is by one author with no agreement statistic; S7, as a lookup, needs none. Signals S2, S3 and S6 are often unverifiable at this sample size and support no conclusion. The price list is one vendor's catalogue and says nothing about market size. E2 and Dcal give a floor, not a rate.
What S7 does not show. S7 records that a cited host can be bought, not that any citation was bought. No claim attaches a transaction to a page, entity, or account, and no entity is named as a manipulator. Confounding remains: user-publishing platforms take 5.2% of citations in E1 against 34.1% in stratum A, so the composition of the citation surface itself varies with commercial motive. The defensible statement is that commercially motivated queries draw on a citation surface that overlaps an open placement market far more than motive-matched controls do.
What I take from it
For anyone using synthetic manipulation corpora: for each attack class, write down which feature channel carries it and whether that channel is populated for the other classes. Where it is not, a reported zero is a property of the generator. The check is cheap, and once it is done the point cannot be argued afterwards.
For detection practice: popularity rankings are convenient weak supervision, and in this sample they point the wrong way. A supply-side lookup is cheaper, needs no judgment, and here carried an odds ratio of 9.00 against 0.67. Its cost is that it identifies exposure, not manipulation, and it exists only where a placement market can be observed.
Citation
Citing the research rather than this note? Cite the paper: 10.5281/zenodo.23719633. That DOI always resolves to the current version.
If you refer to this note, please cite it as:
@misc{an2026cgep,
title = {Purchasable, Not Obscure},
author = {An, Tao},
year = {2026},
howpublished = {\url{https://tao-hpu.github.io/articles/cgep}},
note = {Personal research notes}
}