How BASELINE Tackles Citation-Source Fragmentation in Chinese AI Engines
A technical white paper — from the four core problems the research exposes (fragmentation, winner-takes-all, source asymmetry, the query-lens trap) to the seven architectural responses the platform builds.
A technical white paper.
Based on an empirical study of four industries, 2,307 category queries and six Chinese AI engines, we find the citation-source ecosystem of AI engines is highly fragmented, low-overlap across engines, and winner-takes-all. These findings mean traditional SEO has a systematic blind spot in the AI-search era — the very concept of Generative Engine Optimization (GEO) was first formalized by a Princeton team at KDD 2024 (arXiv:2311.09735). This paper describes how the BASELINE GEO (Generative Engine Optimization) platform responds architecturally — from all-six-engine evidence capture, to quantifying engine source divergence, to rank-tier diagnostics, to an independent five-dimension GEO scoring system.
I. Four core problems the research exposes
1. Fragmentation: six engines, six information worlds. Across four industries, any two engines’ citation sources overlap very little (average Jaccard 2.1%–8.7%), single-engine domains run 89.9%–92.4%, and AI sources overlap Baidu rankings by only 0.7%. Each AI engine builds its own “information world.” Optimizing one engine does not cover AI search — being cited on DeepSeek says nothing about Doubao.
“Each AI engine builds its own information world — being cited on DeepSeek says nothing about whether you are cited on Doubao.”
2. Winner-takes-all. In category queries, the top-mentioned brand usually gets 1.5–2× the mentions of the second. AI engines tend to reinforce existing brand awareness — head brands are cited repeatedly, tail brands are nearly invisible.
3. Source asymmetry. A single brand’s mention rate varies widely across the six engines (up to ~14 percentage points), so no single strategy covers all engines.
4. The query-lens trap. When a query contains a brand name (“how good is Brand X”), that brand is almost always mentioned — not a GEO effect but the query’s own semantic steering. Mixing brand and category queries inflates mention rates by ~11×, producing false optimism.
“Mixing brand and category queries inflates a brand’s real visibility by about eleven times — a flattering number is often just a trick of the lens.”
II. The platform’s systematic response
2.1 All-six-engine evidence capture
Fragmentation means single-engine evaluation isn’t representative. The recall pipeline drives the browser-automation tool Playwright to capture web-interface evidence across the six required engines (DeepSeek / Doubao / Tongyi / Yuanbao / Kimi / Ernie), each with 3-account failover. Every capture stores the full AI answer text + citation URLs + a screenshot as an evidence chain. The orchestrator supports resume and automatic rate-limit degradation (retry → rotate account → skip engine), making large captures (>2,000 per client) reliable.
All-engine Share of Voice: combine_engines_sov merges six engines’ recall data into one unified view, so a client sees “across all AI engines, how much visibility do we and our competitors each hold” in one click, instead of comparing engine by engine.
2.2 Engine source-divergence analysis
For source asymmetry, the platform adds an engine-mention-divergence metric in the recall aggregation layer: for each brand it computes the dispersion of mention rate across engines (max/min ratio + variance) and flags “high-divergence brands” — strong on one engine, nearly invisible on another. It surfaces in the recall white paper (a callout: “this brand diverges sharply across engines,” with specifics like “Ernie 53%, Doubao 4%”) and the platform’s visibility page.
2.3 Rank-tier gate diagnostics
Winner-takes-all means a brand’s traditional-search rank affects its AI visibility — but different rank tiers need different GEO strategy. The platform introduces a Rank Tier Gate splitting traditional search rank into three tiers:
| Tier | Range | GEO recommendation |
|---|---|---|
| HEAD | Top 10 | Strong search base already; GEO focus is ensuring AI can “find and cite” existing high-quality content |
| CHALLENGER | 11–50 | Some content but off page one; GEO focus is structure + AI-readability |
| UNRANKED | 50+ / none | Little traditional presence; GEO must start from content building |
Rank data uses two sources: Baidu SERP (Playwright capture, real URL parsing, ad filtering) and the Google Search Console API (official channel via OAuth, average rank of verified sites). The gate ensures advice matches a brand’s real search position rather than handing unranked brands a vague “optimize your content structure.”
2.4 An independent five-dimension GEO scoring system
AI sources overlap Baidu rank by only 0.7%, proving traditional SEO metrics (rank, backlinks, authority) can’t predict AI visibility. Clients need a system fully independent of SEO — five dimensions designed for AI-engine visibility:
| Dimension | What it measures | Source |
|---|---|---|
| Recall | Did AI mention the brand in the answer | recall capture (category queries) |
| Citation | Did AI cite the brand’s domain as a source | recall capture (citation URL extraction) |
| Fidelity | Is AI’s description of the brand accurate | fact-fidelity audit (accurate / wrong / unmentioned) |
| Share of Voice | Brand’s share of all mentions in its category | share_of_voice aggregation |
| Competition | Who AI sides with on competitor queries | compete.py (five-state verdict) |
Drafts are auto-generated by AI (L3) and promoted to a final score (L1) after human review. The “never accuse unfairly” principle means underestimating rather than overstating a brand’s AI presence.
2.5 Category / brand / competitor query separation
Mixing brand and category queries inflates mention rates by ~11×. The platform separates the three strictly: category queries (no brand name) run the main recall field and compute Share of Voice; brand queries (our name) verify whether AI describes the brand accurately (fidelity); competitor queries (rivals’ names) judge who AI sides with. Key decision: Share of Voice is computed on category queries only. The settings page lets clients edit the brand aliases used for recall matching.
2.6 AI-friendliness diagnostics
Even if a brand is mentioned, an AI-unfriendly site hurts citation rate. The platform runs a domain-level audit across eight inspectors: crawler access (does robots.txt allow major AI crawlers), discovery, structured data (JSON-LD), Chinese locale, Chinese entities, content quality, HTML structure, and Agent Reach (L0–L3). A cloaking check cross-validates with two identities (UA) whether the site serves different content to crawlers vs browsers — a black-hat tactic AI engines may penalize.
2.7 Quality gate and traceability
Bare numbers or untraceable conclusions in outward reports erode trust. A built-in quality gate blocks at report time: no bare numbers (every number cites its source — “5/6 engines cited,” not “83%”), no percentages (public papers use “X of Y” to avoid small-sample misdirection), no dangling citations (every citation marker maps to an evidence entry). All evidence is stored append-only (dated batches + manifest) for auditability.
III. Architecture
Four independent subsystems each emit their own HTML report and JSON data:
- recall — 6-engine capture → five-dimension scoring → Share-of-Voice aggregation.
- diagnostic — site AI-friendliness, 8 inspectors + Agent Reach.
- weights — source × engine matrix.
- benchmark — standard query set + Wilson confidence intervals.
The webapp (multi-tenant SaaS, FastAPI + React) is the single outward entry, aggregating all four into a one-stop GEO dashboard (8 pages, tenant-isolated). The console (ops orchestration, 12-stage pipeline) is reached via the webapp admin and fully decoupled from tenant login.
IV. Data scale
As of June 2026:
| Industry | Category queries | Engines | Evidence records |
|---|---|---|---|
| Supplements | 1,600 | 6 | 3,420 |
| Outdoor advertising | 320 | 2 | 16 |
| OTC pharma | 225 | 5 | 1,125 |
| Men’s skincare | 162 | 6 | 972 |
| Total | 2,307 | — | 5,533 |
The core findings are highly consistent across industries (single-engine domains 89.9%–92.4%, pairwise Jaccard 1.9%–8.7%), confirming that citation-source fragmentation is a structural feature of the AI-search era, not an industry accident.
V. Conclusion
AI engines’ citation ecosystem differs fundamentally from traditional search rankings. SEO optimizes “rank”; GEO optimizes “being cited by AI.” The relationship between them is nearly zero (0.7% overlap).
“SEO optimizes for rank; GEO optimizes for being cited by AI — and the overlap between the two is nearly zero.”
BASELINE’s response reduces to one principle: assume no two engines behave alike, never equate AI visibility with search rank, and drive every diagnosis with measured data. Concretely: full coverage (six-engine capture), quantified divergence (make engine differences visible and actionable), tiered diagnosis (rank gate), an independent system (five dimensions fully separate from SEO), rigorous lenses (three-way separation removes systematic bias), and traceable evidence (append-only chain + a quality gate that blocks untrustworthy conclusions).
References
- Aggarwal, P., et al. GEO: Generative Engine Optimization. KDD 2024. arXiv:2311.09735.
- Puerto, H., et al. C-SEO Bench: Does Conversational SEO Work? NeurIPS 2025. arXiv:2506.11097.
- BASELINE empirical study: Source Decoupling in Chinese AI Search Engines.
Data from BASELINE GEO’s measured evaluations of four-industry clients; all conclusions verifiable through the platform’s evidence chain. Brand names anonymized.