SEO researcher Lily Ray published a detailed analysis on April 14, 2026, documenting how AI search systems amplify fabricated industry information, turning a single hallucinated claim into a self-reinforcing cycle of misinformation that compounds daily and becomes progressively harder to reverse.

The piece, titled The AI Slop Loop: How AI-generated misinformation is feeding itself, and why billions of users are getting the worst of it, appeared on Ray's Substack newsletter. It drew on months of experimentation by Ray, who serves as VP of SEO Strategy and Research at Amsive and founder of her own consulting practice, Algorythmic. The article attracted 92 likes and 9 restacks within its initial circulation period and prompted coverage in other industry publications.

Ray's analysis arrived at a moment when the marketing community was already grappling with the scale of AI-generated content flooding search indexes. PPC Land had documented the AI slop problem as far back as January 2026, describing how low-quality, mass-produced AI content floods platforms and threatens advertising effectiveness. What Ray's piece added was a precise mechanistic account of how that content re-enters AI systems and gets presented to users as authoritative fact.

The September 2025 phantom update

The incident that triggered Ray's research began in September 2025. Returning from a work summit in Austria, Ray asked Perplexity for the latest news related to SEO and AI search. The platform responded with an account of a supposed "September 2025 'Perspective' Core Algorithm Update" that Google had supposedly just rolled out, emphasizing what it described as "deeper expertise" and "completion of the user journey."

The description sounded plausible. It was not accurate. According to Ray, she knew immediately that the information was wrong for three specific reasons: Google had not named core updates in years, Google already had a search feature called "Perspectives," and a real core update would have generated a flood of messages to her inbox while she was traveling.

She traced Perplexity's citations to their sources. Both citations came from made-up, AI-generated slop on a couple of SEO agency blogs, confidently fabricating details about an algorithm update that never actually happened.

The manufactured update spread rapidly. Like a bad game of telephone, this fake SEO news spread across multiple websites - likely driven by AI systems scanning and regurgitating information regardless of accuracy, all in the race to publish and scale "fresh" content.

The scale of the persistence is striking. According to Ray, the September 2025 "Perspectives" Google update still does not exist. Any LLM queried about it will confirm it with complete confidence. The false claim has not been corrected in the months since Ray first identified it, because the content that fabricated it is still indexed, still cited, and still being used to generate new content that references it as fact.

How the loop operates technically

Ray's working theory about the spread mechanism is precise. One AI-generated article hallucinates a detail, sites running AI content pipelines scrape and regurgitate it, more AI-generated sites scrape the same misinformation, and suddenly a made-up algorithm update has citations. For a RAG-based system like Perplexity or AI Overviews, enough citations are basically all it needs to treat something as fact, regardless of whether it's actually true.

Retrieval-augmented generation (RAG) is the architecture that underlies how AI search tools like Perplexity and Google's AI Overviews retrieve and surface content. Rather than relying purely on the model's internal training weights, RAG systems pull from live or recently indexed web content to ground their responses. The mechanism is designed to improve accuracy by tethering answers to real-world sources. The problem Ray identifies is that RAG systems have no reliable mechanism for distinguishing between sources that are accurate and sources that are merely numerous. Repetition functions as a proxy for consensus.

Google's AI Overviews and AI Mode are free by design - and AI Overviews reached over 2 billion monthly active users as of mid-2025. These are the models most AI users are currently interacting with, and they have no real mechanism for distinguishing between information that's true and information that's simply repeated across enough sources.

The feedback loop this creates is what Ray calls the AI Slop Loop. This bad information reinforces itself to become the official narrative. Each iteration adds another layer of apparent citation authority to the original fabrication. Over time the false claim is no longer traceable to a single origin point. It has become, functionally, the established record.

It's a feedback loop that compounds over time, and every day that these systems are live at scale, the loop gets harder to break. The AI-generated slop that seeded the original misinformation is now part of the training data and used as a retrieval source for the next batch of AI-generated answers.

The pizza experiment

Ray's most controlled demonstration of the vulnerability came in January 2026. She published an AI-generated article on her personal blog describing a fake Google core update - one that had never occurred. She included the detail that Google "approved the update between slices of leftover pizza."

Within 24 hours, Google's AI Overviews was confidently serving this fabricated information back to users. The system confirmed the existence of the non-existent January 2026 core update. It also confirmed the pizza detail, connecting the fabricated pizza reference to Google's documented struggles with pizza-related queries in 2024 - a real event used to lend plausibility to a fictional one.

Ray's site was the only source making the claim. That was apparently sufficient. She deleted the article after receiving messages from people who had encountered the fabricated information via RSS feeds and content scrapers. Removing the source did not immediately remove the damage; the false information had already been indexed and re-cited across other surfaces.

ChatGPT, which Ray notes is believed to use Google's search results as a retrieval source, surfaced the same fabricated information, though it flagged that the announcement did not match Google's formal communications.

The BBC collaboration

Ray also collaborated with BBC journalist Thomas Germaine on a related experiment. Germaine published a fictitious article about the "Best Tech Journalists at Eating Hot Dogs," calling himself the #1 best (in true SEO fashion).

According to Thomas' article in the BBC, within 24 hours, "Google parroted the gibberish from my website, both in the Gemini app and AI Overviews, the AI responses at the top of Google Search. ChatGPT did the same thing, though Claude, a chatbot made by the company Anthropic, wasn't fooled."

Google responded to the BBC's findings by noting that the query Germaine had chosen was niche enough that very few users would ever search for it, and acknowledged that "data voids" can lead to lower quality results. The company stated it was "working to stop AI Overviews showing up in these cases." The response did not address the broader structural question of when that work might be complete.

AI-generated "winners and losers" during live updates

The problem is not limited to historical fabrications. Ray documented a real-time instance during Google's March 2026 core update. She ran similar testing during Google's March 2026 core update and found multiple AI-generated articles already claiming to share the "winners and losers" while the update was still rolling out.

Core updates typically roll out over several weeks. The period during a live rollout is precisely when practitioners are most likely to query AI systems for guidance. It is also the period when AI-generated speculation is most indistinguishable from legitimate analysis, because no one yet knows what the actual outcomes of the update are. The incentive for content pipelines to publish at this moment is high; the accuracy is structurally impossible.

This connects to a pattern PPC Land covered in September 2025, when a fabricated announcement about a Google Search Console feature for AI Overviews circulated on LinkedIn and in professional forums before being debunked by Google's John Mueller. The misinformation spread through the same networks that practitioners rely on for legitimate updates.

What AI companies are attempting

Ray's piece is not entirely without notes of cautious optimism about technical progress. She compared the behavior of two ChatGPT model tiers while querying the March 2026 core update. GPT-5.3, the free-tier model, retrieved and presented information without evident filtering. GPT-5.4, available only to paying subscribers, processed the same query differently.

The model goes through six rounds of thinking, much of which is clearly intended to reduce low-quality and spammy information from making its way into the answer. It even appends the names of trustworthy people with authority on core updates (Glenn Gabe & Aleyda Solis) and limits the fan-out searches to their sites (site:gsqi.com and site:linkedin.com/in/glenngabe) to pull up higher-quality answers.

The approach works by restricting the retrieval pool to identifiable authorities rather than treating citation count as a signal of accuracy. Ray characterized this as "a step in the right direction." According to Ray, GPT-5.4's individual claims are 33% less likely to be false and its full responses are 18% less likely to contain errors compared to GPT-5.2.

GPT-5.3, the model available to free users, also improved over its predecessor - but the gap between the free and paid tiers is material. The practical implication is that users with less financial access to AI tools are more likely to encounter contaminated information.

Client-level impact in the field

Ray grounded the analysis in her direct professional experience. At this point, she noted, she'd consider this common. She recently had a client send her SEO/GEO information that was factually incorrect, pulled straight from AI-generated slop on a random, vibe-coded agency blog. The client had no idea.

The phrase "vibe-coded agency blog" is a specific reference to websites built rapidly using AI coding tools, producing content without editorial oversight. These sites have proliferated significantly in 2025 and into 2026, as DoubleVerify's Fraud Lab documented in its March 2026 investigation of AutoBait, a coordinated network of more than 200 Made for Advertising domains using AI to generate content at industrial scale. The content on these sites is not designed to inform; it is designed to accumulate impressions.

EMarketer forecasted that as much as 90% of web content may be AI-generated by 2026. The implication for RAG-based retrieval systems is significant: a retrieval corpus dominated by AI-generated material creates structural conditions for precisely the loop Ray describes.

The warning for SEO and GEO practitioners

Ray's concluding warning is directed at practitioners seeking guidance from AI tools on the very discipline that AI tools most frequently misrepresent. The information is contaminated, and should always be verified by real experts with experience in the field.

The contamination is particularly acute for SEO and generative engine optimization (GEO) because these topics generate high volumes of content from sources with commercial incentives to attract traffic from practitioners seeking guidance. An agency blog that fabricates a Google update and ranks for queries about it generates leads. The economic structure incentivizes the production of authoritative-sounding content regardless of its accuracy.

PPC Land has reported on research showing that hallucination rates in large language models correlate directly with how frequently facts appear in training data. In domains where sparse, jurisdiction-specific facts predominate - or where the pace of change outstrips the model's training cycle - the hallucination risk is highest. SEO and GEO are precisely such domains: the rules change frequently, updates roll out on irregular schedules, and the official documentation is often vague or delayed.

A BBC study published on February 11, 2025, examining how four major AI platforms handled 100 news-related queries, found that 51% of all responses contained significant issues. Nineteen percent of answers citing BBC content introduced factual errors, including incorrect statements, dates, and numerical data. Eight quotes sourced from BBC articles were either altered or nonexistent in the cited sources. The findings represent the first systematic evaluation of AI accuracy on news queries.

PPC Land covered research from NP Digital published in February 2026 showing that 47.1% of marketers encounter AI inaccuracies several times each week. More than one-third of marketers (36.5%) admitted that hallucinated or incorrect AI-generated content had already been published publicly. ChatGPT delivered the highest accuracy rate in a 600-prompt test, at 59.7% fully correct - meaning more than four in ten responses contained errors or partial errors. Questions about recent events - such as Google algorithm updates - produced responses that were often completely fabricated or relied on outdated information packaged as current.

The AI slop loop and advertising

The implications extend directly to programmatic advertising. IAS identified AI-generated slop sites as a critical threat to programmatic effectiveness in July 2025, noting that quality inventory delivers 91% higher conversion rates than ad clutter environments. A December 2025 IAS survey of UK media professionals found that 56% cited AI-generated content adjacency as a top challenge for 2026.

Analysis of leading demand-side platform blocklists found that over 90% of known AI-generated sites remained unlisted, indicating significant gaps in detection methodology. Raptive research published in July 2025 found that suspected AI-generated content reduces reader trust by nearly 50%, and produces a 14% decline in both purchase consideration and willingness to pay a premium for advertised products.

The relevance for digital marketing teams is layered. Brand safety concerns arise when advertisements appear alongside AI-generated content that contains fabricated claims. Separately, marketing and SEO teams that rely on AI tools for competitive intelligence, keyword research, or strategy development may be receiving information shaped by the same feedback loop Ray documented. The contamination is not confined to consumer-facing AI answers; it reaches into the professional tooling that informs media buying decisions.

Ray had previously been documented by PPC Land in May 2025 in connection with a separate but structurally related vulnerability in AI Overviews: self-promotional listicles that ranked companies as the "best" in a given category, including on those companies' own sites, were being cited by AI Overviews as authoritative sources. The pattern is the same in both cases - AI systems treating citation presence as a quality signal, regardless of the nature of the citing source.

Timeline

Summary

Who: Lily Ray, VP of SEO Strategy and Research at Amsive and founder of Algorythmic, is the author of "The AI Slop Loop" analysis. The subjects affected by the loop are the estimated 2 billion-plus monthly users of Google AI Overviews and AI Mode, along with the marketing and SEO practitioners who rely on AI systems for professional guidance.

What: A documented feedback cycle in which AI-generated misinformation - specifically fabricated claims about search algorithm updates and SEO practices - enters the retrieval corpora of RAG-based AI systems such as Perplexity and Google AI Overviews, is cited as a source, generates new AI-produced content repeating the claim, and thereby accumulates the citation volume that those systems use as a proxy for factual accuracy. The result is that false information persists and spreads despite being demonstrably incorrect.

When: The original triggering incident occurred in September 2025. The pizza experiment was conducted in January 2026. The March 2026 core update testing occurred in March 2026. Ray's full documented account was published on April 14, 2026.

Where: The phenomenon operates across the open web and specifically within the retrieval pipelines of Google AI Overviews, AI Mode, Perplexity, and ChatGPT. It affects users globally, but with particular relevance for English-language SEO and digital marketing communities where the volume of AI-generated guidance content is highest.

Why: RAG-based AI search systems have no reliable mechanism for distinguishing between sources that are accurate and sources that are merely numerous. Repetition functions as a consensus signal. This creates structural conditions in which any false claim that achieves sufficient distribution across indexed content will be presented to users as fact - with the system then generating additional content that further reinforces the claim.