Gemini, Claude, Perplexity: Divergent AI Strategies

Marked Divergences in AI Citations

A comprehensive study conducted by Yext, a specialist in online visibility, has highlighted significant divergences in how major artificial intelligence models select their sources to respond to user queries. This analysis, which focused on 17.2 million citations generated in the fourth quarter of 2025, examined four AI models: Gemini, Claude, SearchGPT (ChatGPT), and Perplexity. The aim was to understand how these models choose their sources and how this impacts the online visibility of brands.

Citation Logics of the Four Models

Yext categorized the sources cited by these AIs based on the degree of control exerted by brands, dividing the citations into four distinct categories: Total, Partial, Limited, and None. Gemini stands out with a high proportion of citations from sources under total control, such as official websites, brand blogs, and press spaces. This trend is attributed to its anchoring in Google's ranking system, meaning that to enhance visibility in Gemini, it is crucial to maintain strong performance in traditional SEO, particularly regarding E-E-A-T criteria.

In contrast, Perplexity presents a more stable citation profile across sectors, with a share of proprietary sources ranging from 37% to 50%. This model operates as a full-fledged search engine, triggering a search in its own index for each query, which, according to Yext, favors stability over a highly contextual adaptation.

Claude, on the other hand, is characterized by a significantly higher reliance on user-generated content, classified in the "Limited" category. In the food and beverage sector, this share reaches 24.35%, compared to only 2.57% for Gemini. The study highlights that citations falling under the "Limited Control" category are present at levels two to four times higher than those observed for other models. Yext hypothesizes that the constitutional AI of Anthropic, which underpins Claude, may place greater weight on diverse user-validated sources as signals of relevance.

For SearchGPT, sectoral variability is the most pronounced. In the hospitality sector, the model cites official hotel websites 38.1% of the time, compared to a range of 16.7% to 22.4% for the other three models. The study indicates that this is the most marked divergence observed for a given model within a sector, without identifying the cause with certainty.

Complex Sectoral Gaps to Manage

The study extends a previous analysis published by Yext in October 2025, which examined 6.8 million citations and concluded that 86% of AI citations relied on channels controlled by brands. This new component refines the diagnosis by introducing an additional dimension. The gaps are not only between sectors but also among models. "Citation patterns vary more within sectors than between them," Yext specifies. For example, restaurants and specialty food stores fall under the same sector (food and beverage) but exhibit significantly different citation profiles.

This situation creates a strategic challenge that many companies have yet to fully integrate. The healthcare sector stands out as an exception, with the lowest divergence between models: less than 6 points difference in the share of proprietary sources, compared to 10 to 15 points in most other sectors. Yext sees this as a possible "shared sensitivity to the accuracy of medical information across different model architectures," combined with the very structure of available sources, dominated by large medical directory platforms. In contrast, the food and beverage sector concentrates the most marked gaps, particularly regarding the share of user-generated content.

Companies that view AI search as a single channel to optimize rely on an assumption that current data does not support. For marketing and SEO professionals, this means that a brand can achieve high visibility in Gemini, which values official sites, and remain almost invisible in Claude, which relies more on reviews and social content. "Companies that consider AI search as a single channel to optimize are relying on an assumption that current data does not support," concludes the study. However, Yext notes that this data represents a snapshot of the fourth quarter of 2025 and that frequent model updates could alter these patterns.