Knowledge Graph

Google announced the Knowledge Graph on 16 May 2012 with more than 500 million objects and more than 3.5 billion facts about them^[1], built from public sources including Freebase, Wikipedia and the CIA World Factbook^[5]. Official counts since have been sparse: 70 billion facts by mid-2016^[6], then over 500 billion facts about five billion entities in May 2020^[2]. A 2024 estimate by Kalicube, published in Search Engine Land, put Google's Knowledge Vault at over 1,600 billion facts on 54 billion entities; it is an outside estimate, not a Google figure^[7].

Freebase was one of the public sources the Knowledge Graph drew on at launch^[5]. It held more than 3 billion facts about almost 50 million entities when Google turned it read-only on 31 March 2015^[8]. The migration of that content to Wikidata mapped 4.56 million items, 9.5% of Freebase topics^[9]. Wikidata now holds over 123 million items^[10]. Wikidata and Wikipedia are open descriptions of an entity that anyone can edit.

Google's guidance says AI Overviews and AI Mode may use a query fan-out technique, issuing multiple related searches across subtopics and data sources, and that no special schema.org structured data needs to be added for a page to appear in them^[11].

An analysis of 680 million citations collected between August 2024 and June 2025 found Wikipedia to be ChatGPT's most-cited source at 7.8% of total citations, against 0.6% for Wikipedia in Google AI Overviews^[12]. A 2026 study of 167,551 grounded citations covering 128 brands across 12 markets and 13 languages found Wikipedia the most-cited domain in 11 of 12 languages^[3], and that 85.7% of citations pointed at sites the brand does not own^[4]. The descriptions AI systems draw on for brand mentions therefore sit mostly outside the brand's own domain^[4].

Google's Knowledge Graph Search API returns only individual matching entities rather than graphs of interconnected entities, is not offered for production-critical use, and directs new users to Cloud Enterprise Knowledge Graph^[13]. AI visibility measurement therefore works backwards from what the systems output: knowledge panels, AI citation sources and the answers themselves.