New Audit Reveals Hidden Gap Between AI Search Citations and Rankings

When Google’s AI Overviews answer a question, most users never scroll past the tidy summary that appears at the top of the page. The cited sources tucked beneath that answer are, for millions of people, the new front page of the internet. But a large-scale new audit suggests that what the artificial intelligence chooses to cite and what traditional search rankings display are not the same thing, and the divergence follows patterns that are structured, stable, and consequential for who gets seen online.

The study, published in Information Systems Frontiers by Chiang-Yu Cheng of Soochow University in Taiwan, introduces a diagnostic construct called Citation–Ranking Divergence, or CRD, to measure how far AI-generated citations drift from the ranked results that search engines have long served. Rather than asking whether AI answers are accurate in isolation, the research asks a more structural question: does the generative layer of search allocate visibility to sources the way the ranking layer does, or does it quietly re-shuffle the deck?

To find out, Cheng conducted a large-scale, multilingual, two-wave audit of Google AI Overviews. The audit used 1,500 base queries collected at an initial point in time and then re-collected the same queries at a second point, producing 3,000 wave-specific observations across five language–region locales before eligibility filtering. This repeated-measures design matters because a single snapshot cannot distinguish a temporary quirk from a durable pattern. By comparing across time, the study could test whether the relationship between citations and rankings is a passing artifact or a stable feature of how generative search operates.

The headline finding is deliberately counterintuitive. Citations are neither fully detached from search rankings nor simply reflective of them. Instead, the generative layer remains partially anchored to ranked visibility while selectively re-concentrating exposure through a narrower subset of sources. In practical terms, the AI does not ignore the web’s existing hierarchy of authority, but it does not reproduce it faithfully either. Some sources that rank highly in the classic search results never make it into the AI answer, while a more concentrated group of cited sources captures a disproportionate share of user attention.

This concentration effect is where the study’s stakes become clearest. For two decades, the economics of the open web have rested on a rough bargain: publishers produce content, search engines index it, and rankings send traffic back to the creators. Decades of eye-tracking and click research have shown that position matters enormously, with users overwhelmingly favoring top-ranked results. Generative search disrupts that bargain by giving users the answer directly. In that world, a citation inside an AI summary becomes the scarce currency of visibility, and if the allocation of that currency diverges from rankings, the distribution of attention, traffic, and ultimately revenue shifts in ways publishers may not even be able to detect.

One of the study’s most technically interesting contributions is a typology of divergence patterns. The audit found that visibility reallocation does not follow a single uniform logic; instead, it takes recurring forms. Some divergence events appear as a narrowing, in which the AI draws from a smaller pool of sources than the ranking layer offers. Others involve the promotion of sources that were not prominently ranked, effectively injecting new winners into the visibility ecosystem. By classifying these patterns, the study transforms what might look like random AI behavior into something auditable: a set of recognizable signatures that regulators, researchers, and platforms themselves can look for.

The cross-wave analyses strengthen the case that these patterns are reasonably stable over time. That stability is a double-edged finding. On one hand, it suggests the divergence is not mere noise from a system still in flux, which makes systematic auditing feasible and meaningful. On the other hand, it means the reallocation of visibility is likely to persist, compounding over time as cited sources accumulate authority, links, and training-data presence while uncited sources fade further from view. A stable divergence is, in effect, a standing redistribution of the web’s attention economy.

Methodologically, the study positions itself within the growing tradition of algorithmic auditing, in which researchers probe automated systems at scale to expose behaviors their operators may not disclose. What distinguishes this work is its human-centric framing. The author argues that citations in generative search are not merely transparency cues, the little references that make an answer feel trustworthy. They are mechanisms of source visibility allocation, deciding which publishers exist in the eyes of users who never visit the results page. Reconceptualizing citations this way reframes the auditing question: the goal is not only to check whether sources are real and relevant, but to map who wins and loses exposure when an AI mediates the answer.

The multilingual design also carries weight. Generative AI systems are known to perform unevenly across the world’s languages, and an audit confined to English could easily miss divergences that disproportionately affect smaller linguistic communities. By spanning five language–region locales, the study builds CRD into a scalable framework that can be redeployed in different markets and repeated as models are updated. That scalability is precisely what makes the construct useful beyond a single paper: it offers a template for ongoing, longitudinal oversight of systems that change continuously and largely in private.

Cheng situates these findings within the broader debates on platform accountability, algorithmic transparency, and knowledge visibility governance. As regulators, including those implementing the European Union’s AI Act, grapple with how to oversee generative systems, the study offers a concrete, measurable target: the alignment between what an AI cites and what the underlying ranking infrastructure exposes. If that alignment degrades, transparency alone, in the sense of showing users a list of sources, may not be enough to guarantee fair allocation of attention. A citation list that is visible but structurally concentrated could satisfy the letter of transparency while violating its spirit. The audit suggests that meaningful accountability will require looking past the presence of citations to their distribution, and treating the gap between cited and ranked sources as a governance problem in its own right, one that will shape what the public can know and whose voices it can hear in the age of answer-first search.

Subject of Research: A large-scale multilingual audit of Google AI Overviews examining divergence between AI-generated citations and search rankings as a mechanism of source visibility allocation.

Article Title: Human-Centric Auditing of AI-Powered Generative Search: When Citations Diverge from Rankings

Article References: Cheng, C.-Y. (2026). Human-Centric Auditing of AI-Powered Generative Search: When Citations Diverge from Rankings. Information Systems Frontiers. https://doi.org/10.1007/s10796-026-10813-4

Image Credits: AI Generated

DOI: 10.1007/s10796-026-10813-4

Keywords: generative search, Google AI Overviews, citation-ranking divergence, algorithmic auditing, algorithmic transparency, platform accountability, knowledge visibility, visibility allocation, information systems, search engine rankings, AI citations, platform governance