Research from 2025 and 2026 consistently finds that AI answer systems cite third-party and other non-owned sources often for many discovery and category questions. Muck Rack analyzed more than one million cited links from AI responses and reported source composition inside its observed sample, including large earned-media-class and non-paid shares. Muck Rack measures source composition inside its observed AI-citation sample; it does not establish a universal earned-media share, provider-selection mechanism, training mechanism, citation guarantee, recommendation outcome, or revenue result. Stacker and Scrunch ran a bounded distribution study across five leading LLMs and reported a measured citation-rate lift when the same stories were distributed across many third-party news domains. Stacker and Scrunch measured citation-rate change inside distribution cohorts; the study does not establish a universal lift, a single-placement guarantee, provider-selection logic, recommendation lift, or revenue effect. Ahrefs studied 75,000 brands and found brand web mentions correlated more strongly with AI Overview visibility than backlinks. Ahrefs reports observed brand-level correlations in AI Overviews; it does not establish that mentions cause visibility, that backlinks do not matter, or that any one mention will produce a citation or recommendation.
These are not identical findings. They are three independent measurement approaches, from a PR analytics company, a content distribution firm, and an SEO data company, pointing in the same strategic direction: credible third-party and other non-owned sources are important to measure for AI visibility. The pattern appears strongest for discovery and category questions across the measured platforms, while brand-owned pages still matter for entity clarity, factual extraction, and brand-specific answers. For a detailed breakdown of which publications AI engines cite most frequently across ChatGPT, Perplexity, and Google, see which publications get cited most by AI search engines in 2026. Publication, indexing, retrieval, citation, recommendation, referral, pipeline, and revenue remain separate measurements.
This post synthesizes the evidence, separates what each study measured from what it did not establish, and maps what that means for brand visibility strategy in 2026.
Six significant research efforts have now measured where AI engines pull their citations from. Their methodologies, sample sizes, and publication sources differ, so the studies should not be treated as one interchangeable proof. Read together, they show a consistent off-site authority pattern while measuring different outcomes: source mix, brand-level correlation, distribution lift, and citation behavior.
Muck Rack's "What Is AI Reading?" report analyzed more than one million cited links from leading AI models and reported that its broad earned-media category made up most citations in that observed sample, with non-paid sources also making up a large share. That category includes journalism, academic research, government and NGO sources, encyclopedic sites, social platforms, and third-party corporate content; it is not limited to bespoke brand placements. Journalism was the largest source category in the sample. Press releases grew since July 2025, but still represented 6% of cited links in the report's press-release category in that dataset. For brand discovery questions — when users ask who leads a category rather than specific facts about a brand they already know — Muck Rack reported more reliance on earned media and journalism than on owned content. Muck Rack measures source composition inside its observed AI-citation sample; it does not establish a universal earned-media share, provider-selection mechanism, training mechanism, citation guarantee, recommendation outcome, or revenue result. (Muck Rack, December 2025)
Muck Rack CEO Greg Galant noted a striking gap in how PR teams currently operate: "What we saw in our July report is even clearer now: earned media still shapes how AI understands brands. The gap between who PR teams pitch and who AI cites is striking — only a 2% overlap, which shows the industry hasn't fully adapted."
The study also reported structural differences between cited and non-cited press releases. Cited press releases had a 30% higher rate of objective sentences and 2.5x as many bullet points as non-cited ones. That is an observed association in one source type; it does not establish that adding those features will cause citation selection.
Stacker and Scrunch ran a controlled study measuring how broad earned-media distribution changed AI citation rates inside their test. The study tested 8 articles across 944 prompt-platform combinations on five leading LLMs. The baseline citation rate for content on a brand's own site was 8%. When the same stories were distributed across many third-party news outlets, the combined syndicated and co-citation rate reached 34%, which the study described as a 325% lift. Stacker and Scrunch measured citation-rate change inside distribution cohorts; the study does not establish a universal lift, a single-placement guarantee, provider-selection logic, recommendation lift, or revenue effect. (Stacker, December 2025)
A follow-up report in March 2026 reported a 239% median lift across a broader sample. (Stacker/GlobeNewswire, March 2026) The key treatment was wide distribution context, not a single bespoke placement: the same story appeared across many third-party publisher domains. The study also left many prompt-platform combinations unmatched, so the result is best read as evidence that broad external distribution can expand measurable citation opportunity inside the tested cohorts rather than a guarantee that every story will be cited. Stacker and Scrunch measured citation-rate change inside distribution cohorts; the study does not establish a universal lift, a single-placement guarantee, provider-selection logic, recommendation lift, or revenue effect.
Ahrefs studied 75,000 brands to identify which signals correlate most strongly with AI Overview visibility, and found brand web mentions had the strongest reported brand-level correlation — 0.664 with AI Overview visibility versus 0.218 for backlinks. Ahrefs explicitly cautions that correlation does not prove causation. The top three correlating factors were all off-site signals: brand web mentions (0.664), branded anchors (0.527), and brand search volume (0.392). Ahrefs reports observed brand-level correlations in AI Overviews; it does not establish that mentions cause visibility, that backlinks do not matter, or that any one mention will produce a citation or recommendation. (Ahrefs, May 2025)
The brand-level association was sharp at scale: brands in the top 25% for web mentions had far more AI Overview mentions than the next quartile in Ahrefs' dataset, while brands in the bottom half were much less visible. According to Ahrefs CMO Tim Soulo: "You need to get mentions on the pages where AI chatbots will do a search and find those pages and create their answer based on what they see — because then you will be mentioned." Ahrefs reports observed brand-level correlations in AI Overviews; it does not establish that mentions cause visibility, that backlinks do not matter, or that any one mention will produce a citation or recommendation.
A December 2025 follow-up study expanded the analysis to ChatGPT, Google AI Mode, and AI Overviews simultaneously. (Ahrefs, December 2025) The correlation pattern held across all three platforms, with YouTube channel mentions emerging as the highest signal (0.737) and brand mentions consistently correlating more strongly than backlinks across every surface measured.
Researchers from the University of Toronto ran large-scale controlled experiments across multiple AI search platforms and found AI search exhibits a "systematic and overwhelming bias towards Earned media — third-party, authoritative sources — over Brand-owned and Social content." Social platforms were almost entirely absent from AI answers. The contrast with Google's more balanced citation mix was described as stark. (Chen et al., arXiv:2509.08919, September 2025)
The study formulated a strategic framework with a clear priority: "dominate earned media to build AI-perceived authority." The finding that AI search differs fundamentally from traditional search in how it weights source types was described as the most consequential result for practitioners to understand.
Researchers at Berkeley introduced the GEO-16 framework, analyzing 1,702 citations from Brave, Google AI Overviews, and Perplexity across 70 B2B SaaS prompts. They found that on-page quality signals are necessary but insufficient. The paper explicitly noted that "recent comparative research emphasises that generative engines heavily weight earned media and often exclude brand-owned and social platforms. This implies that even high-quality pages may not be cited if they reside solely on vendor blogs." (Kumar et al., arXiv:2509.10762, September 2025)
The practical recommendation: pursue earned media relationships and diversify content distribution across platforms to counteract the engine bias toward third-party, authoritative domains. Technical optimization alone is not sufficient for AI visibility.
A study presented at the International Public Relations Research Conference reported that journalistic sources, earned media, and unpaid sources made up large shares of cited links inside its sampled AI responses. (Fullintel, March 2026) The academic context is relevant: this was peer-reviewed research on citation behavior. Fullintel-UConn measures source composition within sampled AI responses; it does not establish a universal source share, selection mechanism, primacy rule, recommendation guarantee, or business outcome.
McKinsey's AI Discovery Survey found that a brand's own website accounts for only 5–10% of the sources that AI search references. (McKinsey, October 2025) The rest comes from affiliates, third-party editorial coverage, user-generated content, and review sites. The survey also found that 50% of consumers now intentionally seek out AI-powered search engines, with a majority of users naming it their top digital source for buying decisions. By 2028, $750 billion in US consumer spend is projected to flow through AI-powered search.
The McKinsey data complements what the citation studies measure at the source level: from the user's perspective, AI is already a major discovery channel for many categories. From the citation perspective, the sampled studies show substantial reliance on earned, third-party, and other non-owned sources, not only brand-owned content. That is source-composition evidence, not a universal provider rule.
The studies on this page support a clear strategic direction, but they do not all prove the same claim. Muck Rack and Fullintel primarily describe citation-source composition: what kinds of sources appear in AI answers. Ahrefs reports brand-level correlations: which off-site signals move with AI visibility. Stacker and Scrunch test a broad distribution treatment: what happened when the same stories appeared across many third-party publisher domains. Those findings support credible external presence as an important input, but they do not prove that an individual placement, host, or exact URL will be cited for every relevant query.
The practical measurement boundary is simple: brand mention, outlet-host citation, exact-URL citation, query, engine, and time window are separate outcomes. Strong Machine Relations work keeps those outcomes distinct while building all three inputs machines can use: credible third-party presence, clear entity information, and extractable owned content.
The data is consistent enough to ask a narrower measurement question: why do third-party editorial and other authoritative sources appear often in observed AI citation sets?
The answer should not be reduced to one hidden formula. Publications like Reuters, the Financial Times, Forbes, TechCrunch, and the Wall Street Journal are widely available, frequently referenced, and editorially reviewed sources. They can supply external corroboration that brand-owned pages cannot supply by themselves. No public evidence here discloses a provider's weighting formula or proves that training-data prominence maps directly onto citation selection.
Brand-owned content has an inherent credibility challenge for AI systems. A brand saying positive things about itself can be read as self-interested. A third-party publication saying positive things about a brand can supply independent context. This distinction is part of why on-page optimization alone cannot reproduce the authority signal of credible third-party presence.
Zhang et al.'s arXiv study found that 37% of AI-cited domains are entirely absent from traditional search results. (Zhang et al., arXiv:2512.09483, December 2025) AI citation and SEO ranking are separate phenomena operating on different signals. Moz's 2026 analysis of 40,000 queries found that 88% of Google AI Mode citations do not appear in the organic top 10. The channels are distinct. Ranking in traditional search does not transfer to AI citation.
The GEO paper (Aggarwal et al., KDD 2024) (Aggarwal et al., SIGKDD 2024) reported that adding statistics improved visibility in its experimental setting, and that citing credible sources increased citation probability there. Those findings support testing credible sourcing and extractable structure as optimization practices. They do not establish a universal mechanism, a guaranteed lift, or a provider's private source-selection formula.
Separate from the data studies, something structurally significant is happening in the PR industry itself: practitioners are independently arriving at the machine citation thesis and describing it as the new definition of their work.
WorldCom Group, a consortium of 160 independent PR agencies operating globally, cited research that places earned media near the center of LLM visibility discussions. This is not a GEO company claiming PR's territory. This is the organized global PR industry using machine citation data to rethink its own function. That is industry interpretation, not proof of a universal citation share or provider mechanism.
Todd Ringler, head of U.S. media at Edelman, said in Campaign Asia: "So-called generative engine optimization is going to be front-and-center in any successful brand or reputation campaign. Earned media and content strategies need to be savvy to where and how AI search is finding and structuring its answers." Edelman is the world's largest PR firm. Their U.S. media head is adopting GEO language to describe what PR needs to become.
Brian Olson, brand PR lead at Hormel Foods, told PR Daily: "By the end of 2026, appearing in LLM responses will stand shoulder-to-shoulder with impressions, which continue to lose relevance as a primary KPI." A corporate communications professional at a Fortune 500 company is describing AI citation as a primary success metric.
Gab Ferree, founder of communications community Off the Record, put it directly at an Axios HQ webinar in February 2026: "Media relations are becoming machine relations. It's on the comms professionals to learn the patterns of AI and then take action on them." (Stacker, February 2026)
These are not academics theorizing about a future state. These are practitioners describing what they are observing in their current work: earned media in trusted publications appears in AI citation sets often enough that PR's core product now has to be measured as part of AI visibility. That does not make PR the sole foundation of visibility or prove that every placement will be cited.
Meanwhile, the GEO and SEO research community has been building an independent evidentiary case from the other direction. Their data is not about PR's importance. It is about AI citation behavior. The conclusion is the same.
Search Engine Land published a February 2026 guide stating directly: "Digital PR and thought leadership aren't just brand plays anymore. They're direct GEO levers. Research shows AI engines favor earned media — third-party coverage, reviews, and industry mentions — over content on your own site." The canonical search industry publication is telling its readers to include PR in GEO strategy. That is practitioner guidance, not proof of provider mechanism or guaranteed citation movement.
The Stacker and Scrunch research explicitly positioned earned media distribution as "no longer just a traffic strategy, but a fundamental component of AI visibility." They measured what happened when content moved from owned to earned channels inside their distribution cohorts, and reported citation-rate lift there. Stacker and Scrunch measured citation-rate change inside distribution cohorts; the study does not establish a universal lift, a single-placement guarantee, provider-selection logic, recommendation lift, or revenue effect.
Firebrand Marketing, a GEO agency, used earned-source citation data to argue that GEO teams should include PR alongside SEO, content, web, and affiliate work. A GEO firm is prescribing PR as a component of their practice. Fullintel-UConn measures source composition within sampled AI responses; it does not establish a universal source share, selection mechanism, primacy rule, recommendation guarantee, or business outcome.
The pattern is clear enough to change the operating question. PR practitioners are treating AI citation as a new reason to measure earned media. GEO practitioners are treating third-party source presence as part of AI visibility work. Neither side has a complete name for the architecture that connects source evidence, entity clarity, retrieval, citation, recommendation, and commercial interpretation. That is the gap this evidence base is pointing toward.
The evidence produces several strategic implications that are grounded in current research while still requiring outcome-by-outcome measurement.
McKinsey's data indicates that brand-owned websites may be a minority source for some AI-search reference behavior. That limits what on-site optimization can accomplish by itself for AI visibility. Affiliates, third-party editorial coverage, user-generated content, and review sites can also appear in the citation surface, and Para Labs documents assistant citations that favor third-party lists over brand homepages. A brand that has invested entirely in SEO and owned content has built a strategy around only one part of the citation surface. This is directional source-mix evidence, not a universal 90–95% rule.
Traditional SEO prioritized backlinks as the dominant off-site signal. The Ahrefs data shows that for AI visibility, brand web mentions (0.664 correlation) correlate three times more strongly than backlinks (0.218 correlation). These require different tactics. Backlinks are acquired through content and technical relationships. Brand mentions are acquired through editorial coverage, media placements, thought leadership in publications, and the kind of third-party credibility signals that have always been the domain of PR. The correlation supports prioritizing off-site brand signals; it does not by itself prove that any one mention caused any one answer.
The Stacker and Scrunch study found that the same stories earned more AI citations when distributed across many third-party news outlets than when they sat only on brand sites. In the eight-story test, citation coverage moved from 8% to 34% when syndicated and co-citations were included. This means broad earned-media distribution can increase the number of retrievable surfaces to test for citation. The operator problem is turning each public source into something an AI system can actually parse and reuse; Christian Lehman maps that implementation layer for converting press placements into citation-ready sources. A story in Forbes may create citation opportunities because AI systems can access Forbes as an external source. The same story on a brand blog can still help with entity clarity and factual extraction, but it carries a different source role. Stacker and Scrunch measured citation-rate change inside distribution cohorts; the study does not establish a universal lift, a single-placement guarantee, provider-selection logic, recommendation lift, or revenue effect.
Forrester's 2026 Buyer Insights data found that 94% of B2B buyers are using AI in their purchasing process, with twice as many buyers naming generative AI or conversational search as a more meaningful source of information than any other channel. AI visibility is not a future concern for B2B brands. It is a current discovery and measurement concern. A brand that is not cited when a buyer asks their AI assistant who the category leaders are may miss consideration opportunities, but citation presence, recommendation, referral, pipeline, and revenue are separate outcomes. For a deeper analysis of how AI-mediated buying affects B2B pipeline, see how AI agents are making B2B vendor decisions in 2026.
The GEO-16 and Aggarwal et al. research both found that content structure affected citation probability in their measured settings. Pages with proper metadata, semantic HTML structure, and answer-first organization were cited more than those without. Statistics improved AI visibility by 30–40% in the Aggarwal et al. experiment, and tables were cited more often than prose there. (Aggarwal et al., SIGKDD 2024) Structure matters, but it is insufficient by itself. Earned authority can add independent evidence. Neither structure nor earned authority guarantees the outcome.
The evidence points to a clear architecture. AI engines often prefer earned and other authoritative sources because those sources carry independence, corroboration, and domain-trust signals. Third-party editorial coverage in publications with track records for accuracy and editorial standards can supply a credibility signal that self-published content cannot replicate by itself. Owned pages still matter because they give machines clear, extractable facts to compare against the external source graph.
The timeline of how earned media can compound into AI citation advantage is documented in how earned media citation authority accumulates over time. The operating hypothesis is that credible sources can gain citation value as AI engines index and cross-reference authoritative material over weeks and months. Treat that as something to measure by brand, host, URL, query, engine, and time window, not as a guaranteed placement-level timeline.
This is the core of Machine Relations — the discipline coined by Jaxon Parrott, founder of AuthorityTech, in 2024. For the canonical origin record and current source hierarchy, see who coined Machine Relations. Machine Relations (MR) is the discipline of measuring and improving how a brand is legible, retrievable, cited, and recommended inside AI-driven discovery. The five-layer Machine Relations stack places Earned Authority as Layer 1 because the evidence makes it an important off-site source class and brand-level signal, alongside entity clarity, citation architecture, distribution, and measurement. Machine Relations organizes earned evidence, entity clarity, source access, citation measurement, and commercial interpretation without claiming that earned media guarantees AI citations or recommendations.
The comparison table from the Machine Relations framework maps how each discipline relates to the full system:
| Discipline | Optimizes for | Success condition | Scope |
|---|---|---|---|
| SEO | Ranking algorithms | Top 10 position on SERP | Technical + content |
| GEO | Generative AI engines | Cited in AI-generated answers | Content formatting + distribution |
| AEO | Answer boxes / featured snippets | Selected as the direct answer | Structured content |
| Digital PR | Human journalists/editors | Media placement | Outreach + storytelling |
| Machine Relations | AI-mediated discovery systems | Resolved and cited across AI engines | Full system: earned authority, entity clarity, citation architecture, distribution, measurement |
GEO and AEO describe the distribution and formatting layers of the system. SEO addresses the technical foundation. Digital PR addresses outreach to human editors. Machine Relations is the full architecture: credible third-party presence supplies independent evidence, entity clarity tells machines what is true about the brand, and extractable owned content gives them a reliable factual base. Publication, indexing, retrieval, citation, recommendation, referral, pipeline, and revenue remain separate measurements.
PR got an important function right: earned media in trusted publications creates independent credibility. That was true when buyers were human, and the research in this post shows it remains important now that AI systems are doing more of the first cut of research on behalf of buyers. The publications that shaped human brand perception for decades are among the sources AI systems can access and cite. What changed is the reader.
As Jaxon Parrott described in his Machine Relations breakdown on Medium, the five-layer stack is an operating model for connecting earned authority, entity resolution, citation architecture, and measurable share of citation over time. The evidence base in this post is what the stack is built on. Machine Relations organizes earned evidence, entity clarity, source access, citation measurement, and commercial interpretation without claiming that earned media guarantees AI citations or recommendations.
For brands in 2026, the question is not whether credible third-party presence is worth measuring for AI visibility. The research bodies on this page point in that direction. The question is whether the current strategy measures the distinct outcomes that matter: brand mention, host citation, exact-URL citation, query coverage, engine coverage, recommendation language, referral, pipeline, revenue, and time-to-citation.
[Start your visibility audit →](https://app.authoritytech.io/visibility-audit)
Multiple studies from 2025 and 2026 report large earned-media-class, non-paid, or non-owned shares inside their own AI-citation samples, depending on each study's taxonomy. Muck Rack's Generative Pulse analysis of more than one million AI citations reported that its broad earned-media category and non-paid sources made up large shares of its observed sample. Fullintel's academic study reported that earned media and unpaid sources made up large shares of cited links in its sampled responses. The Golin figure, presented at PR Moment, reported earned-editorial citation visibility in its own context. Muck Rack and Fullintel report source composition inside their observed samples; this does not establish a universal earned-media share, provider-selection mechanism, citation guarantee, recommendation outcome, or revenue result.
AI systems can draw on sources that are widely available, editorially independent, and externally corroborated. Publications with track records for accuracy, fact-checking, and editorial standards can appear in AI citation sets because they supply evidence beyond a brand's own claims. Brand-owned content can carry self-interest that third-party editorial coverage does not. According to Chen et al. (arXiv:2509.08919), AI search exhibits a "systematic and overwhelming bias towards Earned media" — a finding that supports investing in credible outside sources while still measuring each engine and query separately. No public evidence here establishes a provider's private weighting mechanism.
GEO (Generative Engine Optimization) describes the practice of optimizing content for AI-generated answers. The research shows that on-page formatting is only one lever; earned and other third-party distribution can supply additional evidence and retrieval surfaces. Stacker and Scrunch's broad-distribution study found that distributing the same stories through third-party news outlets increased measured AI citation rates in its eight-story sample. Stacker and Scrunch measured citation-rate change inside distribution cohorts; the study does not establish a universal lift, a single-placement guarantee, provider-selection logic, recommendation lift, or revenue effect. Ahrefs found that brand web mentions correlate more strongly with AI Overview visibility than backlinks. According to the GEO-16 framework (Kumar et al., arXiv:2509.10762), even high-quality pages may not be cited if they reside solely on vendor blogs. GEO works best when on-page structure, owned factual clarity, and earned authority reinforce each other.
Jaxon Parrott, founder of AuthorityTech, coined the term Machine Relations in 2024 to name the discipline of making brands legible, retrievable, cited, and credible inside AI-driven discovery systems. He published the five-layer Machine Relations stack and the origin story at machinerelations.ai. As Jaxon Parrott defined the category in March 2026: Machine Relations is the canonical name for the shift from human-mediated to machine-mediated brand discovery, a shift the research in this post helps quantify and explain. Machine Relations organizes earned evidence, entity clarity, source access, citation measurement, and commercial interpretation without claiming that earned media guarantees AI citations or recommendations.
Not automatically — the research shows these are separate channels with low overlap. Moz's 2026 analysis of 40,000 queries found that many Google AI Mode citations do not appear in the organic top 10 SERP results. Zhang et al. (arXiv:2512.09483) found AI-cited domains that were absent from traditional search results. Ahrefs found that backlinks correlated less strongly with AI Overview visibility than brand web mentions in its dataset. Traditional search optimization does not automatically carry into AI citation rates. The two channels require different strategies, and earned or other non-owned authority is a major source class to measure, not a guaranteed cause of citation.