Optyino.ai Study: AI Cites 6.61 Domains Per Answer on Average
The wide variance in citation counts across AI models means that a single GEO strategy is insufficient; brands must tailor their content approach to whether an AI model references many sources (Google AI Mode) or few (ChatGPT, Claude).
Reporting from 1 sources: ASCII.jp.
Wallabee Co., Ltd., through its GEO platform Optyino.ai, analyzed 7,295 generative AI responses to recommendation-type prompts collected from September 2025 to April 2026. The study found that AI models cited an average of 6.61 domains and 8.57 URLs per answer. The citation rate, meaning the percentage of answers containing at least one citation, was 84.9%. Google AI Mode led with the highest average cited domains at 14.36 per answer, followed by Bing Copilot at 11.97 and Google AI Overview at 9.94. In contrast, gpt-5.2 averaged 3.51 domains and Claude Sonnet 4.5 averaged 3.87, showing up to a 4.1 times difference among models. The survey targeted nine prompts such as "recommended portable power supply" and "recommended laptop." Wallabee noted that for GEO strategies, understanding citation counts per AI service is important because models vary widely in how many sources they reference.
Wallabee conducted a second, related analysis of 5,707 AI responses collected from September 2025 to March 2026, this time examining the search queries the models themselves issue when gathering information for product recommendations. The study targeted 10 prompts including "recommended washing machine," "recommended refrigerator," and "recommended headphones."
ChatGPT 5.2 produced the longest average query word count at 8.16 words, suggesting it searches with more specific conditions. Gemini 2.5 Flash averaged 2.71 words per query but issued a total of 3,892 search queries, indicating it casts a wide net with many short keywords. ChatGPT-mini had the highest average search count at 2.97 times per response, with search behavior varying by product category. Claude Sonnet 4.5 averaged 0.82 searches and 2.51 words per query, the lowest search frequency among the models analyzed.
The survey defined "number of cited domains" as websites cited per answer (multiple pages from the same domain count as one) and "number of cited URLs" as individual web pages cited. Wallabee used average cited domains as the primary indicator to measure how many diverse external sites an AI references. The company noted that citation display methods vary by AI service and that the results are based on API response logs, which may differ from how users see responses on web screens.
Synthesized by Yomimono from the 1 cited source below, including Japanese-language reporting where cited, then editorially reviewed before publishing.