One Query, Seven AI Search Engines, Zero Common Sources

On August 10, 2026, an engineer ran two business-related queries through seven AI search engines: Perplexity, Yandex Neurosearch, ChatGPT, Gemini 2.5 Flash, DeepSeek V3.1, GLM 5.2, and Grok 4.5. The goal was to measure which domains each engine cited in its answer. The results showed that no domain was cited by all seven engines; more than half of the domains were mentioned by only one engine. The maximum overlap was five out of seven for a single domain. Furthermore, the pairwise overlap metric revealed that DeepSeek, GLM, and Gemini returned identical lists because they all rely on the same web-search plugin from a single provider. Engines with their own indexes showed different "world models": Yandex Neurosearch cited aggregators and media, ChatGPT cited catalogs and consulting firms, and Grok cited government portals. The author also described five failure points of the methodology, including a silent parsing bug with GigaChat's inline_data.sources (which is a dict, not a list) and the need to normalize punycode domains. The author emphasized that citation does not equal traffic, that measurements are time-sensitive, and that a single run is not a verdict. The cost of the experiment was 404 rubles for one evening, and each query via relay costs about 1.26 rubles, while a Yandex generative response costs about 5 rubles.