As artificial intelligence increasingly mediates how users discover information across the web, digital marketers and search engine optimization professionals have grown heavily focused on securing citations within AI-generated responses. However, new empirical research suggests that these coveted citations may be far more volatile than previously understood. A comprehensive multi-tool study tracking citation stability over a 20-day period reveals that roughly half or more of all AI citations disappear by the very next day, upending traditional assumptions about visibility in conversational search.
The research initiative, which monitored queries across five major AI platforms including ChatGPT, Google AI Overviews, Google AI Mode, Microsoft Copilot, and Gemini, sought to understand how web citations behave over time. Analysts tracked whether cited websites remained visible on subsequent days, whether missing citations eventually returned, and what underlying page characteristics contributed to long-lasting visibility. To ensure the findings were robust, researchers cross-referenced the primary dataset against broader historical information tracked daily within Ahrefs Brand Radar.
Measuring AI Citation Stability Across Five AI Tools
To map the lifecycle of an AI citation, researchers queried five major AI tools with identical questions every day for a 20-day window. Every website cited on the initial day was rigorously tracked to determine its persistence. If a citation vanished, monitors recorded whether it returned later in the tracking period and how long the absence lasted.
The results demonstrate that citation churn is a systemic reality across the conversational AI landscape. Across all evaluated tools, between 50% and 80% of citations present on day one vanished by the following day. Specifically, platforms like Google AI Mode and ChatGPT saw roughly 60% of their citations disappear overnight, while AI Overviews experienced about a 50% drop-off. Microsoft Copilot and Gemini exhibited even higher volatility, with approximately 80% of citations failing to appear on consecutive days, though this was largely attributed to the frequency with which those tools generated answers devoid of any source links.
When researchers accounted for underlying search behavior, they discovered that changing web queries—often referred to as fan-out searches—only partially explained the instability. Even when ChatGPT repeated the exact same backend searches on consecutive days, the cited websites changed the vast majority of the time. This indicates that a significant portion of citation churn occurs during the synthesis phase, when the AI model independently determines which specific web pages to feature in its response.
Despite this rapid turnover, the study found that a missing citation is rarely gone for good. Roughly three-quarters of citations that disappeared at least once returned during the 20-day observation window, with about half of those returning within three days. Only a small fraction of initially cited websites vanished permanently without returning. Furthermore, the longevity of a citation served as a strong predictor of its resilience. Following a dynamic akin to the Lindy effect, citations that managed to survive for over a week showed a remarkably high probability of persisting or returning after a temporary absence, whereas brand-new citations proved highly unstable.
Untangling Apparent Citation Drops from Algorithmic Shifts
The research emphasizes that a drop in AI citations does not always signal a loss of authority or competitive failure. In many cases, fluctuations are driven by structural quirks in how different conversational models formulate answers. For instance, platforms like Gemini and Copilot frequently deliver responses completely devoid of source citations. When researchers excluded these empty answers from the dataset, the underlying stability metrics for Gemini and Copilot aligned closely with those of ChatGPT and Google’s AI features, showing a standard 60% daily turnover rate among cited domains.
Beyond raw link citations, the study also examined the relationship between website links and direct brand mentions within AI outputs. Conversational models frequently link to a domain without explicitly naming the brand in the conversational text, or conversely, mention a brand by name without providing a clickable source. The data revealed that brand names tend to stick around significantly longer than specific URL citations. While cited URLs frequently rotated on a daily basis, brand identifiers maintained a much more consistent presence, highlighting the importance of tracking link citations and brand mentions as distinct performance indicators.
Page Characteristics and Query Types Linked to Lasting Citations
To determine what separates fleeting citations from enduring ones, analysts compared the attributes of pages that maintained their visibility against those that quickly faded. The findings indicate that traditional search engine performance metrics—specifically Google rankings and organic traffic—exhibit a much stronger correlation with citation stability than simple domain authority or raw backlink counts.
Pages that already ranked well within traditional Google search results and accumulated steady organic traffic tended to retain their AI citations for substantially longer periods. For example, pages cited in AI Overviews for 15 days or more boasted significantly higher estimated monthly organic visits compared to pages that appeared for only a single day. By contrast, traditional metrics like Domain Rating showed negligible impact on citation longevity, suggesting that raw link equity is less important for sustaining AI visibility than proven relevance and search demand.
Query intent and topic category also played a definitive role in how citations behaved. When users submitted queries explicitly naming a brand, such as inquiries regarding specific software capabilities or pricing models, the brand’s own official website remained cited far more consistently than third-party sources. However, this dynamic shifted depending on the nature of the question. Factual inquiries about features and pricing overwhelmingly favored the brand’s primary domain, whereas subjective reviews and comparative queries prompted AI systems to lean more heavily on external publishers, Reddit discussions, and YouTube videos.
Ultimately, the findings suggest that digital marketers should avoid treating AI citations as rigid, reliable key performance indicators. Because citation lists fluctuate constantly due to algorithmic synthesis and model updates, evaluating visibility requires a macro-level perspective that measures consistency over time rather than reacting to daily fluctuations.
