AI Search Is Contaminating Itself: The Retrieval Poisoning Crisis and What Google Click Signals Actually Do

AI search systems are contaminating their own outputs through a real-time retrieval loop that requires no retraining cycle to spread misinformation. An Oumi analysis of 4,326 AI Overview responses found that while 85–91% appear accurate on the surface, 56% of correct answers are ungrounded — the cited sources don't actually support the claims. Separately, DOJ antitrust documents finally clarify how Google actually uses click data through Navboost and RankEmbedBERT.

Together, these findings expose two fundamental misunderstandings in the SEO industry: that AI citations equal trustworthiness, and that clicks directly influence rankings. Neither is true — and the gap between perception and reality is widening.

1. The Retrieval Poisoning Crisis: AI Search Is Eating Itself

Unlike traditional model contamination (which requires retraining over months), RAG-based systems like Google AI Overviews, Perplexity, and ChatGPT fetch live web content and present it as authoritative answers. When that live content is itself AI-generated, hallucinated, or fabricated, the contamination is instantaneous. The retrieval layer is not a filter , it is the infection vector.

This is different from the "model collapse" researchers have warned about. Model collapse is a slow degradation over training cycles. Retrieval poisoning is real-time. A speculative blog post published at 9 AM can be cited as authoritative fact by 10 AM. This active connects to the ghost citation problem , AI systems are citing content without verifying it, and now without even verifying that the citations support the claims.

2. The Numbers: How Bad Is the Contamination?

The Oumi analysis reveals a critical distinction between *surface accuracy* and *grounded accuracy*. A response can sound correct while citing sources that don't actually support the claim. Over half of all "correct" answers fall into this category , they give the illusion of citation-backed authority without the substance. Across 5,380 sources analyzed, Facebook and Reddit ranked as the second and fourth most-cited platforms , neither of which has mechanisms to verify human authorship or factual accuracy.

3. The Mechanism: Why RAG Systems Are the Infection Vector

Two academic papers demonstrate the structural vulnerability. PoisonedRAG (Zou et al., 2024) showed that a small number of crafted passages can control RAG system outputs without compromising the model itself , injecting content into the retrieval corpus is sufficient. BadRAG (Xue et al., 2024) demonstrated semantic backdoors enabling similar manipulation through content designed to trigger specific retrieval patterns.

The practical attack chain works like this: an AI content pipeline generates a speculative article → the article gets indexed within hours → a RAG system fetches it during a user query and cites it → other AI pipelines observe the citation and reference the same content → the fabricated claim becomes "consensus" across multiple AI systems without any human verification.

xAI's Grokipedia exemplifies the endpoint of this trend , an AI-rewritten encyclopedia that bases articles on contaminated web content, including Instagram reels as sources. There is no human responsibility mechanism for correcting errors.

4. The SEO Industry's Role in the Contamination Loop

The irony is acute: the SEO industry is simultaneously the victim and the accelerant of this crisis. When AI Overviews and AI search tools began capturing traffic that previously went to publishers, agencies responded by deploying AI content pipelines at scale. But the content these pipelines generate , speculative algorithm analyses, "best X" roundups, generic how-to articles , became the raw material that other AI systems now cite.

This connects to the ChatGPT citation mechanics research showing that 44% of ChatGPT citations are "best X" listicles , the exact content formats that AI pipelines produce at highest volume, typically structured around self-interested product rankings rather than independent evaluation. Meanwhile, human creators are abandoning the open web as the traffic bargain collapses. The content that *would* provide genuine first-hand expertise is increasingly published behind paywalls, in newsletters, or not at all , leaving the open web to synthetic content that AI systems will continue to ingest and cite. The zero-click survival strategies we covered earlier become even more critical in this context.

---

5. Google Click Signals: What the DOJ Documents Actually Reveal

DOJ antitrust documents from September 2025 cut through persistent myths about how Google uses click data. The key finding: clicks are the lowest-level data point, not a ranking factor. They are processed, aggregated, and transformed before influencing anything.

How Click Data Actually Flows Through Google's Systems

The Click Fraction Formula

A 2006 Google patent describes how individual clicks become aggregate signals:

The smoothing constant S0 is critical: it prevents low-volume queries from being gamed by artificial clicks. Individual click manipulation is diluted by the normalization process. This is not a "more clicks = higher ranking" system , it's a statistical aggregation designed to resist exactly that kind of manipulation.

RankEmbedBERT: Less Data, Better Results

The DOJ documents reveal that RankEmbedBERT is trained on 1/100th the data of its predecessors while producing higher quality search results. This suggests Google has shifted from quantity-dependent approaches to architectures that extract more signal from less data , making the quality of training signals (including click-derived ones) more important than their volume.

6. Google's GEO Job Posting: A Mixed Signal

Google's ads organization posted a "GEO Partner Manager, Performance Solutions" role within its Large Customer Sales team. The listing mentions "Generative Engine Optimization" seven times and references analyzing "Share of Model" , a brand's visibility in AI-generated answers.

This is worth monitoring but not overstating. It represents one hiring signal from Google's advertising sales organization. The practical implication: Google's ads team sees commercial opportunity in the GEO space, even if the search quality team doesn't endorse the plan. The "Share of Model" metric is the most interesting element , if Google develops tooling to measure brand visibility within AI-generated answers, that's a signal that AI answer optimization will eventually become a paid advertising surface, not just an organic discovery channel.

Frequently Asked Questions