Ghost Citations, the AI Slop Loop & What's Driving 69 Million Crawler Visits

Kevin Indig defines a ghost citation as what happens when an AI system uses your content, provides a source link, but never mentions your brand name in the answer text. Analysis of 3,981 domains across 115 prompts, 14 countries, and 4 AI engines finds that 74.9% of domains received citations, but only 38.3% were mentioned by name. Of all citations in the dataset, 61.7% were ghost citations: linked, anonymous, and invisible to the user unless they clicked through to check. Only 13.2% of appearances resulted in both a citation and a mention.

The behaviour varies sharply by AI engine. ChatGPT cites 87% of the time but names brands in just 20.7% of cases. Gemini does the opposite: it names brands in 83.7% of cases but only cites 21.4%. The content type matters too. Comparative and evaluative content, pages that explicitly compare options or recommend one over another, generates brand mentions. Informational content feeds AI engines anonymously. The geographic variation is also notable: mention rates range from 50% in India and Sweden down to 18-22% in Italy, Brazil, and the Netherlands.

Duda analysed 858,457 sites on its platform and recorded 68.9 million AI crawler visits in February 2026. OpenAI accounts for 81% of that total, Anthropic 16.6%, Perplexity 1.8%, and Google 0.6%. The purpose of that crawling has also shifted: 56.9% is now User Fetch, meaning real-time content retrieval to generate live answers, rather than training (28.8%) or content discovery (14.3%). ChatGPT is responsible for nearly all of the User Fetch traffic. The majority of AI crawling is happening to serve active users right now, not to build future models.

The data on what predicts crawling is the most directly actionable part. Sites with a Yext integration have a 97.1% AI crawl rate compared to 58% without. Sites with a Google Business Profile sync have a 92.8% crawl rate versus 58.9% without. Structured data, external integrations, and content depth all correlate with higher crawl frequency. Sites receiving AI crawler visits average 527.7 human sessions compared to 164.9 for non-crawled sites, a 3.2x difference. The direction of causation is not certain, but the association between AI crawling and human traffic is consistent across the dataset.

Lily Ray names and documents the AI Slop Loop: a three-stage cycle where one AI-generated article invents false details, multiple AI content pipelines scrape and republish the misinformation, and enough citations accumulate that retrieval systems treat repetition as consensus. The mechanism matters because AI systems often rank credibility partly by how many sources corroborate a claim. If the sources are all downstream copies of the same fabrication, repetition gets mistaken for verification. Ray uses a March 2026 core update as a concrete example, where AI-generated articles invented nonexistent algorithmic details that then spread across dozens of other sites within days.

The scale data is sobering. Google's AI Overviews have a 91% accuracy rate, but applied to 5 trillion annual searches, that 9% error rate means tens of millions of wrong answers per hour. 56% of technically correct AI Overview responses were ungrounded, meaning the cited sources did not fully support the information provided. With ChatGPT's free tier representing 94% of its 900 million weekly active users, and AI Overviews reaching 2 billion monthly users, the loop is operating at a scale where even small error rates become enormous in absolute terms. Newer models are improving: GPT-5.4 produces 33% fewer false claims than GPT-5.2, but the structural incentive to republish and amplify rather than verify persists.

AI Overviews appear far more often for informational queries than for commercial or transactional ones. Ecommerce and decision-stage searches trigger them much less frequently, which means bottom-of-funnel content is relatively protected from the zero-click effect that has hit top-of-funnel pieces hard. Frunze's recommendation is to shift content allocation to 60-80% bottom- and mid-funnel work, with top-of-funnel content repositioned as topical authority infrastructure, passing link equity to conversion-focused pages rather than driving traffic directly.

The attribution challenge makes this shift harder to measure than it sounds. Bottom-of-funnel content often converts through subsequent direct visits or branded searches after a user discovers the brand through an AI citation. That journey shows up as direct or branded traffic in analytics, hiding the role the AI-cited content played in the decision. Standard attribution models undercount the commercial value of content that earns AI citations even when those citations do not drive a direct click. That means the ROI of decision-stage content is systematically underreported in most analytics setups.

SE Ranking analysed 100,000 keywords across 20 niches in New York to compare the March 2026 and December 2025 core updates. March was substantially more disruptive across every metric. 79.5% of TOP 3 positions changed, up from 66.8% after December. Over 24% of pages that held TOP 10 positions dropped out of the TOP 100 entirely, nearly double December's 14.7%. Nearly 30% of current TOP 3 results ranked outside the TOP 20 before the March update, compared to 13% after December. The update created significantly more movement in both directions than the December update had.

Recovery rates were poor. Of domains losing TOP 10 positions, only 23.5% regained them during the update period. The March update also interacted with the earlier spam update: 82% of domains that dropped after the spam update remained absent after the core update, suggesting that spam-penalised sites are unlikely to recover through core update fluctuations alone. One consistent factor across both updates: domain age. Sites 15 years or older held a stable 57% share of TOP 10 results in both December and March, indicating that established domain authority remained protective even through a high-volatility update.