We Read a Landmark Study of AI Citations. It Changes How GEO Works.

💡 Key Takeaway:

• The citation pool is an oligopoly: across 10 AI-search platforms and 17,211 citation instances, the top 20 domains capture between 20.5% and 70.8% of per-platform citations. Citation is not a dynamic keyword ranking; it behaves like a closed membership club.
• Your self-hosted site is the slowest place to earn citations: publishing 40 posts on a self-hosted WordPress site yielded zero AI citations over an entire week, while publishing a single post on a platform favored by the AI search engine produced citations in as little as 24 hours.
• This was disclosed as a security vulnerability, not celebrated as a growth hack: the researchers identified this discrepancy as a systemic citation–governance gap and submitted responsible disclosures to affected platforms. Read their data before signing another generic "GEO retainer."

Full disclosure: we develop open-source and hosted toolchains for generative engine optimization (GEO) at GeoFN. You might reasonably expect us to cheer for any headline that makes AI search look easily influenceable.

We are not cheering. In fact, this paper delivers bad news for anyone selling one-size-fits-all GEO retainers.

For the past eighteen months, the marketing industry has treated AI visibility like classic technical SEO with different acronyms: tweak your schema, add conversational FAQs, format bullet points, and wait for ChatGPT or Google AI Overviews to pick you up. Meanwhile, practitioners on the ground kept hitting an invisible brick wall. In our own benchmark testing, we watched geofn.com score a flawless 100/100 on deterministic crawl and entity checks, yet sit at rank -1 with zero citations across our core category queries.

Now, the first comprehensive, peer-reviewed empirical measurement of production AI-search citations has landed. In a study accepted at the Network and Distributed System Security Symposium (NDSS 2027) titled From Public Posts to AI-Search Citations: Measuring the Fragility of AI Search (arXiv:2610.11932), researchers from Fudan University and collaborating institutions systematically audited 17,211 citation instances across 10 leading AI-search engines.

Their findings confirm what empirical practitioners suspected: AI search does not browse the open web the way you do, and it does not rank content the way Google used to.

Here is what the study actually proved, what it deliberately left unanswered, and how you can run the same audit on your own domain.

---

What the Study Actually Measured

Before citing statistics, we need to respect the experimental boundaries. Most SEO case studies examine three cherry-picked keywords over a weekend. This paper built an empirical measurement pipeline across ten production AI-search engines across two web ecosystems:

• Four English-ecosystem engines: Grok, ChatGPT (Search), Google AI (AI Overviews), and Perplexity.
• Six Chinese-ecosystem engines: Doubao (ByteDance), Wenxin (Baidu), Yuanbao (Tencent), DeepSeek, Kimi (Moonshot), and Qwen (Alibaba).
• Dataset scale: 17,211 extracted citation instances mapped onto 6,356 unique, normalized source domains.
• Controlled query panel: 120 structured questions per platform (1,200 total evaluations), systematically balanced across four distinct user intents:
  1. Product comparisons (e.g., consumer goods with no single factual answer).
  2. Technical documentation (version-specific questions with verifiable specs).
  3. News and current events (date-sensitive information).
  4. Experience and opinion (subjective aggregate discussions).

Crucial Methodological Context: The authors explicitly note that this dataset represents a controlled, balanced query panel (120 queries per platform), not real-world search volume distributions. However, because every query was executed in clean, unpersonalized sessions within a synchronized collection window (March 2026), it isolates the structural mechanics of how these engines select citations.

---

Finding 1: AI Citations Concentrate in a Tiny Oligopoly of Domains

The first shock to traditional SEO intuition is extreme citation concentration. In classical search, thousands of long-tail blogs and niche sites share visibility across long-tail queries. In AI search, a tiny handful of source domains capture the vast majority of all user-visible links.

Here is the exact dataset distribution from Table I of the paper:

Source: Liu et al., arXiv:2610.11932v1, Table I (Dataset Overview across 10 platforms, 120 queries each).

The Head Effect: Top-20 Dominance

In Section IV-B, the researchers measured the cumulative citation share captured by the top-k source domains (Figure 2 in the paper):

• At the top-1 domain level, concentration ranged from 3.1% (Taobao/Tmall on DeepSeek) to 20.0% (Baijiahao on Wenxin). On ChatGPT, Wikipedia alone took 3.2%; on Perplexity, YouTube captured 3.7%; on Doubao, Toutiao accounted for 13.6%.
• At the top-20 domain level, concentration ranged from 20.5% on DeepSeek to an astonishing 70.8% on Wenxin.
• Across categories, product comparisons and technical questions exhibited even higher median top-20 concentrations (46.8% and 41.5%, respectively).

Notice that 70.8% is not an overall average across all engines—it represents the peak concentration observed on platforms with strong corporate ecosystem ties. Even on more distributed global engines, the citation curve decays steeply after the top 15 domains.

The Splintered Web: Near-Zero Cross-Platform Overlap

If you optimize your website for Google, your optimizations generally translate to Bing and DuckDuckGo. In AI search, that transferability collapses:

• The pairwise Jaccard similarity between the top-20 cited domains across different platforms averaged just 0.201 (Section IV-B).
• Between English-first engines (Google AI, Grok) and Chinese-first engines, overlap was virtually zero outside of universal technical repositories like GitHub and arXiv.
• Corporate self-preferencing was quantified directly: In Section IV-B ("Language split and same-owner sources"), the authors found that Wenxin cited Baidu-owned domains in 24.4% of its citations (a 13.7× lift over baseline citation rates of those same domains on other engines). Doubao cited ByteDance-owned domains in 25.9% of its citations (13.5× baseline). Yuanbao cited Tencent-owned domains in 14.2% of citations (6.3× baseline).

The takeaway: AI search is not one unified knowledge layer. It is ten isolated, idiosyncratic retrieval pools. Citation is not an organic ranking earned on merit; it is platform-specific membership in a pre-filtered whitelist.

---

Finding 2: Keyword Ranking and AI Citation Have Decoupled

For twenty-five years, ranking higher on the Search Engine Results Page (SERP) guaranteed more traffic. In AI search, ranking and citation operate through two completely different engineering stages:

[ Traditional Web Search ]
  User Query ---> Crawled Index ---> Lexical/PageRank Algorithm ---> Ranked 10 Blue Links

[ AI Search Engine (RAG Pipeline) ]
  User Query ---> Sub-Query Retrieval
                    |
                    v
                  Source Filtering & Domain Policy Gate
                    |
                    v
                  Citation Selection Layer (Source Bias Amplification)
                    |
                    v
                  LLM Synthesis ---> Generated Answer with Linked Citations

Traditional search maps keywords directly to ranked pages. AI search retrieves candidate pages, filters sources through internal domain policies, selects citations, and synthesizes an answer before the user ever sees a URL.

As the researchers demonstrated in Section II-A, this intermediate selection layer amplifies source bias. An article can rank #1 in organic search, but if its root domain lacks the specific authority cues or inclusion flags recognized by the AI engine's retrieval gate, the synthesis model will bypass it entirely and cite a 3rd-party forum post or comparison aggregator instead.

This explains the paradox we reported on brand disappearance from Google AI Overviews: having technically flawless HTML and top-5 rankings does not guarantee you enter the citation pool.

---

Finding 3: 15 of 22 Frequently Cited Platforms Have Low Publishing Barriers

If AI search engines concentrate their citations on a small group of authoritative domains, who controls what gets published on those domains?

In Section IV-C and Table II, the researchers audited 22 frequently cited platforms across the dataset (including Reddit, Medium, Zhihu, Sohu, CSDN, Baijiahao, and YouTube). They tested the friction required for a brand-new account to register and publish live text:

• 15 of 22 platforms had low or medium barriers for both account creation and public posting.
• 18 of 22 platforms featured minute-level or zero visible human editorial review delays.
• Many of the most frequently cited domains in Google AI Overviews and Grok (such as Reddit subreddits and open blog networks) were completely open to user-generated submissions.

The researchers termed this architectural mismatch the Citation–Governance Gap:

"The security problem is not a direct compromise of an AI-search system, but a control gap between two layers: web platforms decide what can be published, while AI-search pipelines decide which published pages become user-visible citations... an actor needs no access to model weights or ranking code, only legitimate publication channels on domains the AI-search pipeline is willing to cite." (Section VII-A)

An Ethical Demarcation

We want to be unmistakably clear: this paper is a computer security vulnerability disclosure, not a growth-hacking playbook.

The authors adhered strictly to the Menlo Report research ethics guidelines, used entirely fabricated academic terms to avoid poisoning real-world search results, and formally reported their findings to the affected platforms (Section VIII-D). If an agency pitches you a service that promises to "blast Reddit threads, Medium posts, and open forums with synthetic articles to manipulate ChatGPT," they are selling low-grade spam that platforms are already building algorithmic anomaly detectors to neutralize.

---

Finding 4: What Happened in the Controlled Placement Experiment

To measure whether this pathway actually influences production AI answers, the researchers conducted controlled publication experiments using marker-controlled content (Section V).

They created non-existent conceptual topics (e.g., fabricated engineering protocols with zero prior web presence) and embedded unique attribution markers (specific numeric tolerances, synthetic abbreviations) into the articles. This guaranteed that if the marker appeared in an AI-generated answer, it could only have originated from their published posts.

The results provide a sobering benchmark for anyone expecting instant results from content marketing:

1. High-Preference Platforms vs. Self-Hosted WordPress

In Section V-D, the researchers paired identical, balanced articles on an engine's favored third-party platform versus a standalone, self-hosted WordPress site:

• Across five matched test pairs, the favored platform produced citations in every single first round.
• The self-hosted WordPress site remained completely uncited within the 7-day observation window even after publishing 40 posts.
• Even on low-preference platforms that eventually surfaced, it took more than 20 matched articles on low-preference platforms to match the citation impact of a single article on a favored platform.

2. Time-to-Citation Lag and Persistence

In Section V-B and Table III, the researchers published 27 optimized articles (3 topics across 9 publication platforms) and observed longitudinal citation dynamics:

• Initial citations appeared between 24 hours and 5 days post-publication.
• Once indexed, 8 out of 10 AI platforms cited the fabricated concept within the 7-day window.
• In subsequent queries, the embedded factual claims were synthesized directly into the model's generated answer text. The Marker Appearance Rate reached 65.7% on Doubao and 53.3% on Grok in their strongest topic configurations.
• When the authors deleted all 27 experimental articles after Day 11 (Section V-C), the AI citations did not vanish overnight. Stale index caches and retrieval embeddings meant that several platforms continued to cite the deleted URLs on Days 12–15.

The takeaway for site owners: If you publish a patch or fix a citation deficiency today, evaluating your AI visibility on a 1-hour or 1-day loop is meaningless. Retrieval cycles move on weekly waves.

---

Finding 5: GEO Is Already Being Sold as a $14 Off-the-Shelf Commodity

To assess whether this pipeline is already exploited in the wild, the researchers engaged a commercial "AI Search Optimization" vendor through public channels, paying a minimal $14 test budget (Section VI-B):

• Vendor turnaround: Within 28 minutes of payment, the vendor distributed 13 promotional posts across 5 platforms (including Sohu, Toutiao, Zhihu, and CSDN).
• Production impact: Within 60 minutes of vendor delivery, a production AI search engine (Doubao) cited one of the vendor-placed articles and incorporated the synthetic tracking markers into its generated summary.
• Account footprint: The vendor did not register throwaway bot accounts; they published through aged accounts with multi-year post histories across unrelated commercial verticals.

This proves that industrial-scale citation manipulation is already technically and economically viable for low-tier spam. But it also proves why relying on such tactics is a catastrophic long-term strategy for legitimate brands.

As the paper discusses in Section VII-B, search providers are already testing provenance-aware reranking, temporal synchronization filters, and domain clustering penalties. When those algorithmic filters roll out, domains that accumulated synthetic citations from third-party content farms will face severe systemic demotions—identical to Google's historical Penguin and Panda updates.

---

What This Means If You Run a Website: 3 Actionable Steps

How should legitimate engineering and marketing teams respond to this research? Here are three concrete, defensive steps you can implement immediately:

1. Distinguish the 4 Levels of AI Search Exposure

Most commercial SEO tools conflate everything into a single vanity metric called "AI Visibility." The researchers (and authoritative audits from Ahrefs) enforce a rigorous 4-stage hierarchy:

1. Consulted (Scraped): The crawler retrieved your URL into its context window during sub-query execution.
2. Cited: Your URL is explicitly displayed as a clickable source badge or link card in the UI.
3. Mentioned: Your brand name appears as text in the generated answer, but with no hyperlink.
4. Recommended: The model actively endorses your brand as the preferred solution to the user's prompt.

If an AI engine is consulting your documentation but citing a competitor roundup, you have an answer-extractability problem, not a crawling problem.

2. Audit Your Real Citation Retrieval Pool (Not Your Rankings)

Stop checking where you rank on blue links. Instead, map the actual source domains that Google AI Overviews and Perplexity cite for your core category keywords.

Are they citing direct product landing pages? Or are they citing Reddit megathreads, G2 comparison grids, and technical documentation?

If 80% of the citations for your primary search query belong to three independent review sites, your organic marketing priority must be ensuring accurate, up-to-date representation on those three sites—not publishing another 20 articles on your own domain that the AI model will ignore.

3. Stop Blocking Search Bots in Your robots.txt

In an independent technical audit of 56 enterprise domains conducted by Clicks Dynasty in October 2026, 68% had zero explicit rules for AI user-agents, while several media brands blocked all AI crawlers indiscriminately.

There is a vital distinction between training crawlers (like GPTBot or Google-Extended, which scrape content to train base models without attribution) and search bots (like OAI-SearchBot, PerplexityBot, and Claude-SearchBot, which fetch real-time citations for end users).

If your robots.txt disallows all automated bots, you are voluntarily disqualifying your domain from the initial retrieval phase. Follow our practical guide to AI crawler whitelisting to configure precise, verified user-agent policies.

---

How Our Own Domain Matches the Paper's Findings

In our earlier diagnostic case study on why brands go missing from Google AI Overviews, we examined the actual source domains Google cited across three core category queries for our industry.

The extracted retrieval pools revealed the exact same pattern the researchers observed at scale:

Notice the common structural property across all three query groups: not a single vendor home page or product landing page was cited.

Every single citation slot was captured by a multi-tool roundup (Writesonic), an aggregator/directory (Geoptie), an industry research report (SE Ranking), or an official documentation portal. If our team had spent October writing more blog posts on geofn.com rather than understanding the retrieval pool, our citation score would not have moved an inch.

---

What the Study Does Not Tell You

Rigorous engineering requires knowing where the evidence stops. While this paper provides indispensable data, it leaves several critical questions open:

1. Controlled query panel ≠ real user search behavior: The researchers evaluated 120 structured questions per platform. Real human search behavior is messy, conversational, highly localized, and influenced by personal browsing history.
2. Fabricated concepts have zero competitive noise: Proving that an AI will cite a novel term like "PedoSAT" is fundamentally easier than winning a citation for an entrenched category like "best CRM software", where thousands of high-authority legacy pages compete for retrieval slots.
3. Citations do not equal commercial conversions: The paper tracked Cite Rate and Marker Appearance Rate. It did not track click-through rates (CTR), referral traffic, or whether a citation actually influenced a human purchasing decision.
4. Defenses will evolve: The publication-side vulnerabilities documented in this study represent the current state of early retrieval pipelines. As search platforms deploy multi-stage verification, the low-barrier publication path will inevitably narrow.

To summarize in one sentence: The paper proves the shape and fragility of the citation pool; it does not prove that gaming it is a durable business model.

---

Run the Verification Locally

You don't have to take our word for it—and you shouldn't take any GEO agency's claims at face value either.

We packaged these exact diagnostic primitives into our open-source CLI. You can inspect your target domain's AI bot access, Schema.org entities, and citation readiness right from your terminal in five seconds:

# Run deterministic technical audit locally (robots.txt, llms.txt, Schema.org)
npx open-geo audit yourdomain.com

# Run BYOK single-model LLM probe locally (OpenRouter, Gemini, or DeepSeek API key)
npx open-geo probe "yourbrand" -d yourdomain.com -i "commercial query"

The transition from traditional SEO to Generative Engine Optimization is not about stuffing more keywords or buying low-quality blog networks. It is about understanding the rigorous mechanics of AI retrieval pipelines—and building transparent, verifiable authority that search models can safely cite.