The PR industry's most-repeated statistic — earned media drives 84% of AI citations — is simultaneously true and misleading. Three independent studies measuring the same phenomenon report 25%, 39.5%, and 84%. The gap between those numbers is where the actionable intelligence lives.
Three Studies, Three Numbers, One Problem #
Muck Rack's "What Is AI Reading?" report analyzed 25 million cited links across ChatGPT, Claude, and Gemini responses spanning 17 industries. Finding: earned media accounts for 84% of all AI citations, stable between 82% and 89% across three editions from July 2025 through May 2026.
Meltwater's GenAI Lens examined approximately 5.35 million citations across eight AI engines — ChatGPT, Claude, Gemini, Perplexity, Grok, Google AI Overviews, Google AI Mode, and CoPilot. Finding: earned and news sources account for 39.5% of all citations in April 2026, up from 38.3% in March.
Muck Rack's own earlier Generative Pulse Report, analyzing over one million links across four engines including Perplexity, found earned media at approximately 25% of all LLM citations.
Same company. Same general topic. Different numbers by a factor of three.
A fourth data point adds another layer: 5W Public Relations' 2026 State of AI Search report, citing University of Toronto research, puts the figure at 85.5% and finds that AI engines cite earned media roughly five times more frequently than brand-owned websites. That study also found brands appearing on four or more third-party platforms are 2.8 times more likely to be cited in ChatGPT responses than single-platform brands.
Four studies, four methodologies, and a spread from 25% to 85.5%. The number you get depends entirely on what you count.
Why the Numbers Disagree #
The discrepancy is not a data quality problem. It is a definitional one.
What counts as "earned media" changes everything. Muck Rack's 84% figure uses a broad definition: everything that is not brand-owned content or paid advertising. Under this classification, Reddit threads, Wikipedia articles, G2 reviews, YouTube videos, Quora answers, and industry forum posts all count as "earned media" alongside traditional press coverage. The 84% stat is technically accurate — the vast majority of what AI engines cite is third-party content rather than brand websites.
Meltwater's 39.5% uses a narrower bucket: earned and news sources specifically, separating out social platforms (7.2% and growing), user-generated content, and reference sites. When YouTube, Reddit, and LinkedIn are their own categories rather than "earned media," the number drops by more than half.
Muck Rack's own journalism-specific measurement lands at 25-27% — the share of citations that come from actual editorial press coverage. This is what PR teams directly influence through pitching, relationship building, and media strategy.
The engines measured also shift the result. Meltwater tracked eight engines. Muck Rack's flagship report tracked three. Including Perplexity, Grok, and Google's AI surfaces changes the composition because each engine has distinct citation behaviors. Meltwater's data shows ChatGPT sources 51.1% from earned/news, while Grok pulls 15.1% from social platforms and 40.5% from earned/news. Perplexity and Google AI Mode lean heavily on YouTube — which appears in the top five sources for six of eight models tracked.
The Visibility Gap Inside Earned Media #
Even within the 25-27% of citations that come from journalism, not all press coverage translates to AI visibility.
Connective3 analyzed 3,500 digital PR links across 170 brands and 224 campaigns spanning 2,300 publisher domains. Their finding: 43% of brand mentions were not retained when AI engines summarized the articles containing them. Getting press coverage and being visible to AI are different outcomes.
Their Brand Citation Score research quantified the gap:
| Score Range | AI Survival Rate |
|---|---|
| 0-25 (Low) | 14.72% |
| 25-50 | 29.46% |
| 50-75 | 55.20% |
| 75-100 (High) | 69.95% |
Links with high brand citation scores are 4.75 times more likely to survive AI summarization than low-scoring links. The difference comes down to brand positioning within the article: central placement in headlines, direct expert quotes, strong entity salience, and relevant semantic associations. A brand buried in paragraph eight of a general industry roundup has a 15% chance of making it into an AI response. A brand with headline-level positioning and direct attribution has a 70% chance.
This means the effective share of AI citations that PR teams influence is not 84%, not 25%, but closer to 14-17% — the journalism slice minus the invisible fraction.
The gap between "mentioned in an article" and "cited by an AI engine" matters because the trust signal multiplier is enormous. Research from Seer Interactive found a 75x gap between brands with third-party trust signals and those without — the largest multiplier measured in generative engine optimization research to date. But that multiplier applies to the mentions AI engines actually process, not all mentions in all articles.
The Paywall and Platform Access Problem #
Everything PR's 2026 Visibility Index exposes another layer of the measurement gap: paywall status. At a sample of 40 paywalled publications, ChatGPT and Claude showed no measurable citation lift from paywalled newspaper content. The publications that PR teams consider highest-prestige placements — major national newspapers with hard paywalls — may not be accessible to AI retrieval systems at all.
This creates a disconnect between traditional PR value and AI citation value. A placement in a paywalled Tier-1 outlet that drives significant referral traffic and brand authority in traditional media measurement may contribute nothing to AI visibility. Agility PR Solutions notes that this shift is forcing PR teams to reconsider publication targeting — optimizing for AI-accessible outlets rather than purely for domain authority or audience size.
What the Source-Role Framework Shows #
The earned-versus-owned binary obscures the more useful question: what function does a source serve when an AI engine decides to cite it?
The Machine Relations Index categorizes cited sources by their operational role rather than their ownership structure. Market databases like G2 and Crunchbase serve a different citation function than analyst research from Forbes or journalism from trade publications. User-generated platforms like Reddit serve yet another. Each source role has distinct citation patterns across engines, verticals, and query types.
This role-based taxonomy resolves the measurement debate because it explains why different counting methods produce different numbers. A G2 review page counts as "earned media" under Muck Rack's broad definition but functions as a market database in citation behavior — engines cite it for product comparisons and vendor shortlists, not for the same reasons they cite a press article. Collapsing these into one category makes the 84% number meaningless for operational planning.
What Actually Predicts Citation #
The research synthesis across all three studies points to four factors that predict citation inclusion regardless of how "earned media" is defined:
Structural positioning matters more than placement volume. Connective3's 4.75x multiplier between high and low brand citation scores shows that how a brand appears in a source matters more than how many sources mention it. Muck Rack's data supports this: cited press releases contain roughly twice as many statistics, 30% more action verbs, 2.5 times as many bullet points, and 30% higher rates of objective sentences compared to uncited releases.
Recency compresses the window. Over half of all AI citations reference content published within the prior 11 months, with the highest citation rates for material released within seven days. This means earned media's citation value depreciates faster than its SEO value — a press placement that ranks for months in Google may only drive AI citations for weeks.
Engine-specific behavior defeats universal strategy. ChatGPT cites in 96% of responses with an average of five citations per response. Claude cites in 55% of responses but averages 13 citations when it does. Gemini cites in 82% with eight citations average. A visibility strategy optimized for ChatGPT's frequent-but-shallow citation pattern will underperform on Claude's selective-but-deep pattern.
The journalist-AI overlap is minimal. Muck Rack found only 2% overlap between journalists most frequently pitched by PR professionals and those most frequently cited by AI engines. The publications and writers that PR teams prioritize for traditional media relations are largely different from the sources AI engines select for citations.
The Emerging Measurement Standard: Citation Share #
The PR industry is converging on a new KPI to replace impressions and AVE. Ronn Torossian defines "citation share" as a brand's slice of the answers AI engines return when buyers ask category questions — measured across ChatGPT, Claude, Perplexity, Gemini, and Google AI Overviews. Martech.org's analysis frames it bluntly: AI visibility depends on who writes about your brand, not what you publish yourself.
Citation share is a better metric than aggregate earned media percentage because it is brand-specific, category-specific, and engine-specific. But it still requires disambiguation: a brand's citation share in purchase-intent queries (where market databases dominate) is a fundamentally different measurement than citation share in thought-leadership queries (where journalism and analyst research dominate).
What Operators Should Measure Instead #
The 84% stat answers a question no operator needs answered — whether AI engines cite third-party content more than brand websites. The answer is obviously yes, and knowing it changes nothing about what to do next.
The useful measurements are:
-
Source-role citation share — what percentage of citations in your category come from market databases, analyst research, journalism, user-generated content, and reference sites, measured per engine. This tells you where to invest.
-
Brand citation survival rate — of the articles that mention your brand, what percentage retain that mention when AI engines summarize or cite the source. This is Connective3's contribution and the most actionable metric for PR teams.
-
Citation recency curve — how quickly does a new placement's citation rate decay across each engine. This determines PR cadence and refresh strategy.
-
Engine-specific source preference — which engines favor which source types for which query categories. Meltwater's data showing Grok's 217% month-over-month growth in LinkedIn citations versus ChatGPT's 0.5% social citation rate illustrates why this matters.
FAQ #
Does earned media really drive 84% of AI citations? #
Technically yes, if "earned media" includes Reddit, Wikipedia, YouTube, G2, forums, and every other third-party source. If you mean press coverage and journalism specifically, the number is 25-27% based on Muck Rack's own data. Meltwater's eight-engine measurement puts earned and news sources at 39.5%.
Why do different studies report such different numbers for earned media citations? #
The studies define "earned media" differently, measure different AI engines, and use different sample sizes. Muck Rack's 84% includes all non-owned, non-paid content across three engines. Meltwater's 39.5% separates social and user-generated content across eight engines. The narrower the definition and the more engines measured, the lower the number.
What percentage of PR coverage actually shows up in AI responses? #
Connective3's research across 3,500 PR links and 170 brands found that 43% of brand mentions are not retained when AI summarizes the articles. Links with strong brand positioning (headline placement, expert quotes, semantic relevance) are 4.75 times more likely to survive AI summarization than weakly positioned mentions.
How should brands measure earned media's impact on AI visibility? #
Move beyond the earned-versus-owned binary. Measure source-role citation share per engine, brand citation survival rate across your press coverage, citation recency curves by placement type, and engine-specific source preferences. These four metrics provide actionable intelligence that aggregate percentages cannot.
Last updated: July 17, 2026