Tag: Earned media

About the paper

What is AI Reading? by Muck Rack’s Generative Pulse examines which sources generative AI systems cite when answering realistic consumer prompts.

The report is a modelling/data-pack style citation analysis based on a large prompt set submitted to ChatGPT, Claude and Gemini, with more than 25 million cited links analysed across multiple industries; the geographic scope is not clearly specified in the report.

Length: 43 pages

More information / download:
https://generativepulse.ai/report/

Core Insights

1. What is the central finding of the report about the sources AI systems cite?

The report’s central argument is that generative AI citations are overwhelmingly shaped by non-paid, earned, third-party sources rather than paid media or advertising. About 99% of links cited by AI come from non-paid media, while paid and advertorial content accounts for only 0.3% of all citations. Press releases account for 1.1%.

Within that non-paid universe, journalism remains a major foundation of AI visibility. The report finds that about 27% of all links cited by AI are journalistic. This is framed as a consistent pattern across Muck Rack’s previous studies, where journalism has generally accounted for 20–30% of AI citations.

However, journalism is not the only important source category. The report’s pie charts on pages 5–7 show a broader mix: corporate blogs and content at 24%, aggregators and encyclopedic sources at 17.4%, owned media at 13.7%, government/NGO sources at 8.6%, academic/research sources at 4%, social/UGC at 2.9%, tech platforms at 0.9%, press releases at 1.1%, and paid/advertorial at 0.3%.

The implication is clear: visibility in AI-generated answers is not mainly bought through advertising. It is earned through the kinds of sources AI systems treat as credible, relevant or useful: journalism, reference sources, third-party content, government data, academic material, user-generated platforms and some owned content.

2. How do ChatGPT, Claude and Gemini differ in their citation behaviour?

The report argues that each AI provider has a distinct citation ecosystem. ChatGPT, Claude and Gemini do not simply cite the same sources at different volumes; they appear to rely on meaningfully different source environments.

ChatGPT is described as a “near-universal citer”. It includes citations in 96% of responses, but averages only about five sources per response. Claude is more selective: only 55% of its responses include citations, but when it does cite, it averages 13 sources per response. Gemini sits between the two, citing in 82% of responses and averaging eight sources per response.

The differences become even sharper when looking at the actual domains. ChatGPT’s top cited domain is Wikipedia, followed by Axios, YouTube, Kiplinger and Forbes. Claude’s top domain is PubMed Central, followed by Wikipedia, Quora, ScienceDirect and NerdWallet. Gemini’s top cited domain is Reddit, followed by YouTube, Quora, Wikipedia and NIH.

The report’s interpretation is that these are “three effectively separate information environments”. For PR and communications teams, this matters because AI visibility cannot be reduced to a single generic “AI search” strategy. A brand, journalist, outlet or source may matter a great deal in one model and be nearly invisible in another.

3. What determines whether journalism, press releases or other content gets cited?

The report finds that the type of question asked strongly shapes the type of source cited. Industry trend queries are especially likely to draw on journalism: 46% of industry trend responses cite journalistic sources, more than twice the rate for how-to and comparative queries.

By contrast, how-to queries are less journalism-driven. The report says AI tends to rely more on reference content and brand-owned material when users ask how to do something. Comparative evaluation and best-of queries also behave differently, often pulling in review content, platforms, maps, rankings or consumer advice sources depending on the category.

Press releases are most likely to appear in industry trend responses, but even there they remain a relatively small part of the citation mix. Around 1.16% of industry trend citations are press releases, compared with 0.33% for best-of queries, 0.27% for risk/due diligence, 0.25% for problem/discovery, 0.13% for comparative evaluation and 0.09% for how-to.

Recency also matters, especially for journalism. Among journalism citations with known publish dates, 57% were published within the previous 12 months. The report notes a sharp peak in the first month after publication, followed by a decline through month six and then a long tail. Older articles still matter, but the bias toward recent coverage is clear.

4. Which sources, platforms and outlets stand out most in the report?

Several sources stand out because they behave differently from the broader pattern.

Wikipedia is especially important for ChatGPT and Claude. It appears among the top three cited domains in 12 of 17 industries for ChatGPT and 8 of 17 industries for Claude. For Gemini, it appears in the top three in only 3 of 17 industries, because Gemini is more strongly shaped by Reddit, brand domains and Q&A platforms such as Quora.

Reddit is particularly important for Gemini. The report says Reddit is Gemini’s single most-cited domain, accounting for about 2.4% of all Gemini citations. By contrast, ChatGPT cited Reddit only 16 times and Claude cited it zero times in the study.

YouTube also differs by provider. It accounts for about 2.1% of Gemini citations and about 2.0% of ChatGPT citations, while Claude returned zero YouTube citations across the study. Claude does cite other video platforms such as TikTok and Vimeo, but not at the same level.

Among journalism outlets, the standout is Axios. The report says journalism citations are spread across more than 20,000 distinct outlets, with no single publication generally dominating. Axios is the exception: it appears in ChatGPT’s top three cited domains across 13 of 17 industries. The Associated Press appears in the top three for one industry, while The New York Times and Reuters do not appear in the top three most cited sources for any industry in the report.

5. What are the main implications for PR and communications teams?

The report’s main implication is that AI visibility is increasingly connected to earned authority across multiple information environments. Traditional media relations still matters, but not in a simple “get mentioned in top-tier media” way.

First, journalism remains important because it accounts for about 27% of all AI citations and is especially influential for industry trend queries. For brands that want to shape how AI explains what is happening in a sector, credible media coverage appears to be particularly valuable.

Second, the report suggests that citation strategy must be model-specific. A communications team optimising for ChatGPT would pay close attention to Wikipedia, Axios, YouTube and sector-specific sources. A team concerned with Gemini would need to understand Reddit, YouTube, Quora and other user-generated or community-driven environments. For Claude, academic, research, personal finance and reference-style sources appear more prominent.

Third, owned media still matters, but it is not enough on its own. Owned media accounts for 13.7% of citations, while corporate blogs and content account for 24%. The distinction in the report is important: third-party corporate/blog content is categorised as earned when it is not owned by the company or product targeted in the query, while first-party corporate/blog content is owned media. This suggests that corporate content can influence AI, but third-party validation remains highly significant.

Fourth, communications teams need to think beyond classic media lists. Depending on the sector and query type, AI may cite Google Maps, TripAdvisor, Reddit, Quora, YouTube, PubMed Central, NIH, government websites, academic databases, review sites, ranking platforms or trade publications. The relevant source ecosystem changes by industry.

Finally, the report implies that AI visibility is dynamic. The methodology section explicitly notes that generative AI systems are rapidly evolving and opaque, and that observed behaviours may shift as models are updated or retrained. So the findings should be treated as a snapshot of AI citation behaviour in May 2026, not as a permanent rulebook.

About the paper

The paper is a mixed-methods, proprietary analysis from Muck Rack on how web-enabled generative AI models cite sources in response to realistic prompts.

The report says it analysed more than 1,000,000 links generated by Gemini, Perplexity, Claude and ChatGPT between July and December 2025, across a large prompt set spanning multiple industries; the number of prompts and the geographic scope of the data are not clearly specified in the report.

Length: 35 pages

More information / download:
https://generativepulse.ai/report/

Core Insights

1. What is the report fundamentally trying to understand about AI citation behaviour?

The report is trying to map what kinds of sources generative AI systems cite, how often they cite them, and what seems to influence those choices. It is not mainly a consumer study or a survey of users. Instead, it is an observational analysis of AI outputs and their linked citations across multiple models. Muck Rack frames this as relevant to Generative Engine Optimisation (GEO) and to PR and communications teams that want to understand how brands surface in AI-generated answers.

At a practical level, the report studies citation behaviour across several dimensions: source type, recency, authority, industry specificity, and the balance between earned and owned media. It also emphasises that model behaviour is unstable and can change as systems are updated, so the findings should be treated as a snapshot rather than a fixed rulebook.

The deeper argument is that AI visibility is shaped by media ecosystems, not just by a brand’s own website. For broad discovery-style prompts, the models appear to lean heavily on earned media and reputable third-party coverage; for narrower factual questions, owned channels become more important. That distinction is one of the report’s most important strategic takeaways.

2. What does the report find about the kinds of sources AI cites most often?

The central finding is that non-paid and earned media still dominate AI citations. The report states that about 94% of links cited by AI are non-paid media and that 82% of cited links come from earned media. Journalistic sources alone account for about a quarter of all citations. On the chart on page 5, the mix is shown as 24.7% journalistic, 24.5% third-party corporate/blog earned, 14.4% aggregators/encyclopaedic sources, 12% first-party corporate/blog owned, 7.3% academic/research, 6.7% government/NGO, 6% press release, and 4.3% social/UGC.

That matters because it pushes against the idea that AI answers are driven mainly by brand-owned content. The report argues that traditional journalism remains highly influential, with journalistic citations holding fairly steady in the 20–30% range even though they fell slightly versus July.

The findings also suggest that models do not all rely on the same handful of publications. The page 9 table shows different “top media outlets cited” for Claude, ChatGPT and Gemini, with only limited overlap. Reuters appears prominently for ChatGPT, while Gemini’s list includes Forbes, Investopedia and PC Magazine, and Claude’s includes U.S. News & World Report, Nature and Yahoo Finance. The report’s interpretation is that AI models pull from a diverse set of high-authority outlets, not one universally dominant list.

3. According to the report, what makes content more likely to be cited by AI?

The report highlights three main factors: authority, freshness, and relevance to the query domain. It explicitly says outlet authority matters, both broadly, through high-domain-authority publishers such as Reuters, and more narrowly, through specialist sources for specialist topics. It also says niche outlets remain important for industry-specific queries.

Freshness is a major part of the story. The report says that half of all citations are to material published in the last 11 months, and that the highest citation rate occurs within the first seven days after publication. For Claude and ChatGPT, roughly 4% of all citations come from the last week, 5% from the last two weeks, and 8% from the last month. That suggests AI systems disproportionately reward timely coverage, especially soon after publication.

The report also claims that press release structure matters. On page 15, the infographic says cited press releases have about twice as many statistics, 30% more action verbs, 2.5 times as many bullet points, mention more unique companies/products, and have a 30% higher rate of objective sentences than non-cited releases. That implies that structure and information density may affect machine readability and citation likelihood, although the report does not provide a full technical methodology for how those textual features were measured.

4. What changes and trends does the report identify between July and December 2025?

One of the report’s strongest messages is that AI citation behaviour is not stable. It repeatedly says that models are constantly tuning their citation mix, and page 19 gives concrete examples: ChatGPT reduced its reliance on Wikipedia, Gemini briefly spiked on YouTube in November, and Reuters gradually increased in importance across models.

Several category-level shifts stand out. First, press release citations increased materially. The report says press releases overall rose from 1.2% to about 6% of citations from July to December, while direct citations to PR Newswire, Business Wire and GlobeNewswire rose from 0.2% in July to 1% in December, which it describes as a fivefold increase.

Second, third-party corporate/blog citations declined. The report says this category fell by 35%, dropping from 37% to 24%, and links that decline partly to reduced reliance on management consulting content.

Third, the report notes shifts in sector-specific citation patterns. For example, education queries are said to lean more towards .gov and .org sources than in July; healthcare is dominated by NGO and government sources, with Gemini as a partial exception because it uses YouTube; travel queries now mix in more Reddit and YouTube for some models; and technology queries appear to cite fewer unique outlets than before.

5. What are the report’s main implications for PR, media relations and brand visibility in AI?

The report’s practical conclusion is that earned media drives discovery in AI, while owned media matters mainly for fact-finding questions. It gives examples: broad prompts such as what coffee maker to buy are influenced by reputable earned coverage, whereas specific questions such as warranty details tend to pull from owned documentation. The report explicitly says owned content is important, but only for certain question types.

For PR teams, this means AI visibility is not just an SEO or website issue. It is a media strategy issue. The report argues that for a given brand, much of the relevant AI citation coverage comes from a relatively small set of outlets: page 16 says 50% of a brand’s coverage can come from only 20 outlets. But it also warns there is no universal magic list; each brand has its own citation-driving mix.

A particularly provocative implication appears on page 17: the overlap between the journalists most pitched by brands and those most cited by AI is only 2% on average. If that figure holds up, it suggests many media relations habits are poorly aligned with how AI systems actually construct answers. In other words, the report implies that PR professionals may need to rethink not just message discipline, but who they target, what formats they produce, and how quickly they publish.

One note of caution: the report is useful, but the methodology remains somewhat thin in places. It clearly states the models used, timeframe, and the volume of links analysed, but it does not clearly specify the number of prompts, the exact sampling design, or the geographic boundaries of the dataset. So the strategic patterns are valuable, but some claims should be treated as directional rather than definitive.