1、What Is AI Reading?Executive Summary Citations affect responses:Simply enabling or disabling the ability for AI to search the web drastically modifies responses,indicating that the systems are truly basing their responses on the cited works.Journalism and earned media are important drivers:More than
2、 95%of links cited by AI are non-paid coverage.Of those,over 27%of links are journalistic content.Recency wins:Particularly in OpenAI models,fresh content,especially on topical,opinion-based,or event-driven queries,is prioritized.Query framing changes sources:Advice-seeking or opinion-based prompts
3、trigger more dynamic citations,while encyclopedic queries tend to fall back on older,static training data.Outlet authority matters:High-domain authority outlets such as Reuters,Axios and Financial Times are frequently cited,but not consistently.Sources vary by industry too with only about 15%of sour
4、ces appearing in the top 10 across multiple industries.MethodologyWe analyzed 1,000,000+links from AI responsesThis study explores how modern generative AI systems cite sources in response to realistic user prompts.Our objective was to quantify and characterize the nature of AI-generated citations a
5、cross different use cases and vendor models.This includes their frequency,source types,and the prominence of earned and owned media.To accomplish this,we constructed a large,diverse prompt set and executed it across several web-enabled language models,followed by systematic analysis of the responses
6、 and the cited links.The prompts span a variety of industries and subject matter.Sometimes they specifically mention companies by name,sometimes they do not.The following specific models were used to execute the queries,during the month of July 2025:Chat GPT(both 4o and 4o-mini),Gemini(flash and pro