Gemini
Gemini is Google's family of large language models and the consumer app built on them. It is also the model layer under AI Overviews and AI Mode inside Google Search[1]. The Gemini app passed 1 billion monthly active users on 11 August 2026[2], and Gemini 3.5 Flash became the default model in AI Mode worldwide at Google I/O 2026[1].
Scale inside search
AI Mode passed 1 billion monthly users roughly a year after its debut, with query volume more than doubling every quarter since launch[3]. Measured across 55,393 trending queries sampled over 40 days in 2026, an AI Overview appeared on 13.7% of queries overall and on 64.7% of question-form queries[4].
Grounding with Google Search
Grounding is the mechanism that ties a Gemini request to live results: one request may trigger one or more queries to Google Search, so the model answers from current pages instead of training data alone[5]. It reached the Gemini API and Google AI Studio on 31 October 2024[6]. Gemini 3.x models carry 5,000 free grounded search requests a month, shared across the 3.x family[7]. Past the free allocation, grounding on the 2.5 models is billed at 35 dollars per 1,000 grounded prompts[8].
How sources are chosen
Citation is not a mirror of the ranking. Nearly 30% of the domains cited in AI Overviews appear nowhere in Google's own first page of organic results for the same query, which points to a selection mechanism distinct from the ranking algorithm[9]. Selection also tracks intent: in a study of 31 SaaS keywords, educational articles supplied 70.2% of citations on informational queries, while listicles and buying guides led on commercial ones[10]. The same study saw Google swap at least one cited source between repeated runs for 52% of keywords[11].
Fidelity of the answer to its sources
The answer text and the cited page do not always agree. Splitting AI Overview answers into 98,020 atomic claims, one measurement study found 11.0% unsupported by the pages cited beside them, mostly through omission rather than contradiction[12]. The SaaS keyword study put direct support at 64.7% of the claims it examined[11]. Publishers absorb the cost of the arrangement: well over half of AI Overview cited pages carry display advertising, so an answer that satisfies the query removes the click that paid for the page[13]. See zero-click search and E-E-A-T.
The links themselves fail as well. In the Tow Center's audit of eight AI search tools, more than half of the responses from Gemini and Grok 3 cited fabricated or broken URLs[14]. That is a separate defect from an unsupported claim, and it is invisible to a reader who does not follow the link. Citation accuracy covers the audit and its method.
References (14)
- Google Search's I/O 2026 updates: AI agents and more Archive
- Google's Gemini app hits 1 billion monthly active users Archive
- Google Search's I/O 2026 updates: AI agents and more Archive
- Measuring Google AI Overviews: Activation, Source Quality, Claim Fidelity, and Publisher Impact Archive
- Gemini Developer API pricing Archive
- Google's Gemini API and AI Studio get grounding with Google Search Archive
- Gemini Developer API pricing Archive
- Gemini Developer API pricing Archive
- Measuring Google AI Overviews: Activation, Source Quality, Claim Fidelity, and Publisher Impact Archive
- How Does AI Overview Pick its Sources? (Study) Archive
- How Does AI Overview Pick its Sources? (Study) Archive
- Measuring Google AI Overviews: Activation, Source Quality, Claim Fidelity, and Publisher Impact Archive
- Measuring Google AI Overviews: Activation, Source Quality, Claim Fidelity, and Publisher Impact Archive
- AI Search Has a Citation Problem
Last updated 2026-09-04. Written and maintained by Baseline Labs.