LLMs are now shaping how the public perceives government performance, and the picture they present looks nothing like the news we all consume. That's the central finding of Medianet's Opening the Black Box Phase 2 report, a study of how large language models like ChatGPT, Gemini and Claude cite and represent brands and public institutions.
Phase 2 analysed 957 LLM responses and hand-coded 11,127 citations, with a focus on the Victorian Government's presence in AI responses tied to the My Melbourne Road campaign. Four findings stand out for anyone working in public sector communications.
Medianet Managing Director Amrita Sidhu frames the shift as a governance issue, not a technology trend. "LLMs are a stakeholder that will need their own type of management," she said. "They are an emerging reputational channel that needs to be measured, monitored and managed. Our research is the first step in building the framework to do that."
Traditional media coverage of the Victorian Government was 50% unfavourable. LLM responses on the same subject matter were 14% unfavourable and 43% favourable.
It is the same government in the same period with two opposite verdicts, depending on which channel someone asks.
Jacquie, our Head of Insights, traced the gap back to sourcing. "In Phase 2, we drew a connection between the sentiment towards Vic Government and the content information contained in LLM responses," she said. "These were similarly lifted from Vic Government websites."
When journalists write, they apply scrutiny and editorial judgement. AI models pull from what's published and structured. Government websites, fact sheets and policy pages are written to inform, not to critique, so the sentiment skews accordingly. A department's own site is doing more reputational work in AI environments than most communications teams have accounted for.
Government sources made up 44% of all citations in the study, ahead of corporate content at 23% and traditional media at 16%.
Individual Victorian Government domains carried the weight. vic.gov.au was cited at 5.7%. tac.vic.gov.au 3.8% , transport.vic.gov.au 2.3%, bigbuild.vic.gov.au 2.2%, premier.vic.gov.au 2.2%, and parliament.vic.gov.au 2%.
Combined, these six domains account for close to a fifth of every citation recorded in Phase 2. Victorian Government sources generated more than 3,500 citations, much more than any single corporate brand in the study.
AI models treat such structured, publicly accessible government content as fact. Agencies that publish detailed, well-organised information get cited on the strength of it.
News coverage of a government subject typically runs its course in two to four weeks. AI models operate on a different clock. Once an issue is indexed, it can stay in a model's responses indefinitely.
Adam Palmer, Media Analyst at Medianet, in our recent webinar, pointed to the Royal Commission into Antisemitism and Social Cohesion as a live case study. University leaders were absorbing a rough week of headlines, but he argued the headlines weren't the real risk. "What won't pass is the RC's official report," he said. "That means being very careful with the statements and written submissions they make to the RC, because that is what will influence the official record."
The same principle holds for any government agency facing an inquiry or review. Formal submissions and the public record carry more weight in AI environments than daily headlines do. The data backs this up. The 2019 Banking Royal Commission was still surfacing in LLM responses in 2025, six years on. A December 2024 ASIC ruling that drew heavy traditional media coverage hadn't reached LLM responses at all by the time of analysis. Some issues take six months or longer to appear in AI responses. Once they do, they don't leave.
The dominance of owned content raises a harder question. If government material drives most AI citations, and AI models soften controversy by default, does public accountability suffer?
A critical story about government spending or policy failure can reach readers within days through traditional media. The same story might not surface in AI responses for six months, if it surfaces at all. In that window, someone asking an AI assistant about the same issue gets a materially different answer to someone reading the news. Public sector reporting specifically, where scrutiny of spending and policy decisions is the point, the distance adds up.
It's compounded by a separate structural issue. Much of Australia's highest-quality journalism sits behind paywalls or crawl restrictions, making it largely invisible to AI models regardless of its accuracy or public interest value.
The research also flagged a pattern worth watching. In 137 responses, AI models cited RACV content verbatim without naming RACV at all. For a campaign built around public education, that may not matter as the message still resonates, but for a government program designed to build recognition of a specific initiative, funding announcement or agency, it can be a measurement problem. The department can do the work and get none of the visibility.
Four priorities that follow directly from the data.
Medianet tracks this through its own LLM monitoring tool, part of the Medianet Labs Suite.
Medianet Labs is an exclusive access for clients where we launch some of our most AI-forward tools that empower media monitoring. It currently has our LLM Monitoring tool and PR Optimiser live, with new tools such as Hansard Monitoring and Media Database Chatbot along the way. Learn more about it here.
The findings are from phase 2 of our latest pioneering research into generative citation analysis - Opening the Black Box.
Read more about the Findings from the report.
Download the full report here.