Why GEO Doesn't Work Without Earned Media

The GEO playbook has a problem. Not a minor one. A structural one that multiple independent research teams have now documented.

Generative Engine Optimization promised a path to AI visibility that looked reassuringly familiar: optimize your content, add structured data, insert statistics, write in FAQ format. The same logic that made technical SEO a discipline with a clear checklist. Except when researchers actually tested whether these tactics produce AI citations, the results came back uncomfortable.

Most C-SEO methods -- the content rewriting techniques at the heart of GEO strategy -- are largely ineffective at improving how often AI systems cite a document. Some make things worse.

This finding comes from Parameter Lab and the University of Darmstadt, published in 2025, in what is currently the most rigorous benchmark test of GEO content tactics to date. Separate research from the University of Toronto, Ahrefs, Muck Rack, and a growing number of independent studies all converge on the same conclusion: AI engines are not primarily selecting content based on how that content is structured. They are selecting based on where it lives and who published it.

Earned media -- third-party coverage in publications AI engines already trust -- is not a complement to GEO strategy. It is the foundation without which GEO produces very little. The distinction matters because earned authority is the layer that all other AI visibility tactics depend on.

Key Takeaways

• The C-SEO Bench study (Parameter Lab and University of Darmstadt, 2025) found that most content optimization techniques marketed as GEO tactics are "largely ineffective" -- some produce negative effects on AI citation rates
• The University of Toronto's large-scale GEO study (arXiv:2509.08919) found AI search shows "a systematic and overwhelming bias towards Earned media" over Brand-owned and Social content
• Ahrefs' study of 75,000 brands found brand web mentions correlate more strongly with AI Overview brand visibility than backlinks (0.664 vs 0.218 correlation, not proof of cause)
• Muck Rack's What Is AI Reading? (May 2026), covering more than 25 million cited links, found 84% came from earned media, the sources brands neither own nor pay for — a category spanning journalism, third-party corporate content, Wikipedia, Reddit, government and academic sources, with journalism alone at about 27%.
• Pages with strong on-page GEO scores but no third-party earned media presence still underperform in AI citation compared to those with earned coverage from trusted publications
• For founders and marketing executives: earned media coverage in Tier 1 publications is not a "nice to have" alongside GEO. It is the primary driver of AI citation authority.

What does academic research say about GEO content optimization?

The C-SEO Bench benchmark was published in June 2025 by researchers at Parameter Lab (Technical University of Darmstadt), the University of Mannheim, and associated institutions. It is the first study designed to evaluate whether GEO-style content rewriting actually improves citation rates in conversational AI search engines like ChatGPT, Perplexity, and Google AI Mode.

The methodology was tight: 15 commonly recommended C-SEO methods, tested across question-answering and product recommendation tasks, across multiple domains, with varying numbers of competitors adopting the same tactics simultaneously.

The conclusion: most current C-SEO methods are not only largely ineffective but also "frequently have a negative impact on document ranking, which is opposite to what is expected."

The exception -- the approach that did work -- was improving actual source authority in the LLM context window. The approach that helped was not rewriting content for machines, but being the kind of source machines already trust.

The full C-SEO Bench benchmark is available on arXiv for practitioners who want to verify the methodology.

This finding does not mean content structure is irrelevant. Separate research from Wrodium Research (the GEO-16 framework, published September 2025) found that technical page quality signals -- recency metadata, semantic HTML, structured data -- do correlate with citation probability, with an odds ratio of 4.2 for the overall GEO score. The GEO-16 paper documented this across 1,702 citations from Brave, Google AIO, and Perplexity, auditing 1,100 unique URLs.

GEO-16 also reports that the pages Perplexity cited had a mean GEO score of 0.300 out of 1.0, against 0.727 for Brave and 0.687 for Google AI Overviews, so Perplexity cited many pages that score low by technical GEO standards. The study did not test source authority.

What neither paper disputes is what the University of Toronto research (Chen et al., September 2025 arXiv preprint) found in a parallel large-scale study: AI search exhibits "a systematic and overwhelming bias towards Earned media (third-party, authoritative sources) over Brand-owned and Social content, a stark contrast to Google's more balanced mix." The bias toward external publications is not marginal. It is the dominant structural fact of how AI engines select what to cite.

What does earned media do that GEO tactics cannot?

GEO tactics operate on the assumption that AI engines select content primarily based on how that content is written: its structure, formatting, keyword density, FAQ architecture. This model was borrowed from SEO, where the causal link between on-page factors and ranking is real and documented.

AI search engines work differently at the source selection layer.

When a user asks ChatGPT, Perplexity, or Google AI Mode a question about your industry, those systems pull from sources they have already evaluated for external credibility. That evaluation is thought to happen largely before query time rather than by reading your content then: during training and index-building, by assessing where and how often your brand appears across trusted third-party publications. No study cited here measures when it happens.

Ahrefs measured a related correlation in a 2025 study of 75,000 brands. Brand web mentions -- mentions of the brand name anywhere across the open web, linked or not -- correlated with AI Overview brand visibility at 0.664. Traditional SEO backlinks correlated at 0.218. That is a correlation in one study, not proof of cause, but it points away from links and toward third-party coverage.

Tim Soulo, CMO of Ahrefs, said: "You just need to see where your competitors are mentioned, where you are mentioned, where your industry is mentioned. And you have to get mentions there -- because then if the AI chatbot would do a search and find those pages and create their answer based on what they see on those pages, you will be mentioned."

Muck Rack's What Is AI Reading? analysis of more than 25 million cited links confirmed the source profile: 84% came from earned media, the sources brands neither own nor pay for — journalism, third-party corporate content, Wikipedia, Reddit, government and academic sources together — with journalism alone accounting for about 27%. Paid and advertorial content accounted for a fraction of a percent.

Tactic	What it optimizes	Primary AI citation signal?
On-page FAQ sections	Content extractability	Weak: depends on source authority first
Statistics and citations in content	Content credibility signals	Moderate: but only if the source is already trusted
Schema / structured data	Technical machine-readability	Weak: Perplexity mean GEO score of cited pages = 0.30
Backlinks	Domain authority in search	Weaker: 0.218 correlation with AI Overview brand visibility (Ahrefs)
Earned media in Tier 1 publications	Third-party trust signals	Stronger: 0.664 correlation with AI Overview brand visibility (Ahrefs); 84% of AI-cited links from earned media, the sources brands neither own nor pay for, spanning journalism, third-party corporate content, Wikipedia, Reddit, government and academic sources, with journalism alone at about 27% (Muck Rack, May 2026).

What does the evidence show about earned media and AI citation rates?

The quantitative evidence across several studies points the same way. Stacker and Scrunch published a first-look study in December 2025 of eight stories republished across third-party news outlets: across 944 prompt-platform combinations on five LLMs, about 34% of answers cited some version of the stories, against about 8% citing only the brand's original, which Stacker reports as a 325% lift. Most of it came from republished copies being cited rather than the brand's own page. Stacker sells that syndication and calls the study limited, so read it as direction, not a rate any program will reproduce. The direction is the one this argument predicts: the citations moved to coverage on outside publishers.

A 2026 study presented at the International Public Relations Research Conference, conducted by Fullintel and researchers at the University of Connecticut, ran 400 prompts across 10 personas on weight-loss-drug questions, with the cited articles collected through Scrunch AI. Third-party news and informational sources accounted for 47% of cited links; corporate, university and health-network sites accounted for 48%. The study is notable because it comes from communications researchers, not GEO practitioners — and because even in a category dominated by institutional publishers, roughly half the citation surface sat outside the brand's control.

The Signal Genesys study, analyzing 179.5 million citation records across six LLM platforms, found Perplexity drives the largest citation volume of any AI platform and is among the fastest to index new content from trusted sources.

Meanwhile, AuthorityTech's practitioner read on earned versus owned citation rates argues that the gap Stacker and Scrunch's eight-story pilot observed between brand-only and republished versions is not bridgeable by optimizing the owned side harder. Adding more structured data, FAQ sections, or schema markup to pages without earned media coverage does not close the gap. The foundation has to be there first. The measured unit is citation-rate change inside the measured distribution cohorts. MR earned-vs-owned does not establish a universal earned-versus-owned multiple, that earned media determines citation or recommendation, or that citation produces pipeline or revenue.

Why do most GEO strategies fail without earned media?

The GEO discipline grew from SEO, and it inherited SEO's mental model: visibility is determined by what happens on your own properties. Pages, site speed, structured data, keyword mapping. The channel changed but the logic felt familiar, so it got adopted without much interrogation.

The problem is that AI engines like ChatGPT, Perplexity, and Google Gemini were not trained to treat your site as the authoritative source on your brand. They were trained on the open web -- and the open web's authoritative signal for any brand, person, or concept is the accumulated body of third-party coverage across credible publications.

This is not a design flaw in AI search. It is the same logic that makes third-party coverage more credible to humans than self-published claims. The signal AI engines learned to trust is the same signal that determines human credibility: independent corroboration from sources with their own editorial standards.

When a brand has extensive Tier 1 coverage in publications like Forbes, TechCrunch, Harvard Business Review, or Reuters, AI engines learn about that brand from sources they have assigned high trust scores. When that same brand invests in technical GEO -- schema, FAQ pages, long-form structured content -- without earned media, they are writing the signal in a language AI engines learned to discount: self-assertion.

How did PR discover the AI citation answer before GEO?

Here is the unusual part of this story. The conclusion that primary academic research and data analytics firms have been building toward in 2025 and 2026 -- that AI citation is driven by earned third-party coverage -- was the operating thesis of the public relations industry for decades. PR's mechanism always was: get independent, credible third parties to say true things about your brand, in publications that carry their own editorial credibility. That is exactly what AI engines reward.

But PR got almost everything else wrong. The retainer model charged whether or not placements happened. Cold-pitch approaches burned journalist relationships faster than they built them. The industry scaled headcount instead of relationships.

On the other side, GEO practitioners built the measurement frameworks, the structured content doctrine, and the citation architecture that makes earned media extractable by AI engines. GEO's technical contribution is real and valuable. But without the foundation of earned media, GEO is structuring content that AI engines will not cite at meaningful rates.

Todd Ringler, head of U.S. media at Edelman, described GEO as "going to be front-and-center in any successful brand or reputation campaign" -- while explicitly framing it as dependent on earned media and content strategies. The world's largest PR firm adopting GEO language while emphasizing PR's foundational role. Both sides arriving at the same structure from different directions.

The WorldCom PR Group, a global network of independently owned PR agencies, stated: "Up to 90% of citations driving brand visibility in LLMs come from earned media, positioning public relations at the center of this transformation." That finding appeared in an analysis designed to make the case for PR's relevance -- but the data they used to make it came from GEO researchers.

PR is proving GEO's thesis. GEO data is proving PR's thesis. Neither has the architecture that connects the two.

What is the Machine Relations framework for AI visibility?

That architecture is what Machine Relations describes -- the full system that Jaxon Parrott coined in 2024 after watching both industries converge on the same structural truth from opposite directions. The discipline of ensuring brands are recognized, cited, and recommended by AI systems that now mediate the first cut of B2B research.

The five-layer Machine Relations stack makes the connection between earned media and GEO explicit. At its core, Machine Relations names the full system of brand discovery in an AI-mediated world -- the parent category that GEO, AEO, and AI SEO are each partial descriptions of:

Layer	What it does	What fails without it
Earned Authority	Tier 1 placements that AI engines index and trust	Everything downstream. You're structuring content no one will cite.
Entity Clarity	Consistent brand identity signals across the web	AI engines cannot confidently resolve who you are
Citation Architecture	Structured content that AI can extract and attribute	Even good coverage doesn't translate into clean citations
Distribution Across Answer Surfaces	Ensuring the brand appears in responses across engines	Coverage exists but never surfaces at query time
Measurement	Tracking share of citation across AI engines	No feedback loop to know what's working

The C-SEO Bench finding -- that most optimization tactics don't work -- describes what happens when practitioners operate at Layers 3 and 4 without Layer 1. Structuring content for extraction before establishing the third-party authority that makes that content worth extracting.

What should founders and CMOs do about GEO and earned media?

Forrester named AI visibility "the defining priority for B2B marketing leaders in 2026," citing a poll of 150 B2B marketers in which 69% said AI visibility is now a top CMO or CEO priority. Forrester's research describes what happens when brands lose visibility in answer engines as "a collapse of visibility that destabilizes the traditional revenue engine." What Forrester doesn't say -- but the research does -- is that the path back runs through earned media, not technical optimization.

If your team is investing in GEO -- and it should be -- the first question is what the earned media foundation looks like. Not the technical GEO audit. Not the FAQ architecture review. The publication audit: where does your brand appear, in which publications, with what attribution quality, and how recently?

For brands with strong earned coverage, GEO optimization -- structured content, schema, FAQ sections, entity markup -- can meaningfully improve how that coverage is extracted, attributed, and surfaced across AI engines. The technical work has real return when the foundation is there.

For brands with thin or no earned media, GEO optimization produces structurally limited results. You can have a perfect GEO-16 score, clean schema, FAQ-formatted pages, and consistently cited statistics, and still not appear in AI responses for your core competitive queries. AI engines are sourcing those responses from publications that have covered your competitors with independent editorial authority.

Search Engine Land's 2026 GEO guide put it directly: "Digital PR and thought leadership are direct GEO levers. Research shows AI engines favor earned media -- third-party coverage, reviews, and industry mentions -- over content on your own site." That is the canonical search industry publication telling its own readership that PR is now a GEO tactic.

The research literature is direct about this. A September 2025 arXiv preprint, not yet refereed, lists dominating earned media to build AI-perceived authority among four needs it emphasizes, alongside engineering content for machine scannability, adopting engine-specific and language-aware strategies, and overcoming big-brand bias.

For founders making resource allocation decisions between earned media investment and technical GEO investment: build the foundation first. Earned media in Tier 1 publications that your prospects and AI engines both read. Editorial relationships, not spray-and-pray pitch volume. Outcomes-based placement, not retainer cycles that produce no citations.

GEO tells you how to structure content so AI can extract it. Earned media tells AI engines your content is worth extracting in the first place. Christian Lehman, Chief Growth Officer of AuthorityTech, has written on how this plays out in practice: the companies that treat GEO and earned media as sequential rather than parallel consistently underperform against those that build both simultaneously.

Start your visibility audit →

FAQ

Does this mean on-page GEO optimization is useless?

No. The GEO-16 framework found a meaningful quality correlation with citation probability, and the research from Tian et al. (arXiv:2603.09296) showed a 40% relative improvement in citation rates from targeted content repairs. GEO-16 found a correlation and did not test how page quality interacts with source authority; our view is that content repairs apply to sources already being considered, and that without earned media there may be few citations to improve.

Which GEO tactics have the most evidence behind them?

The GEO-16 findings point to recency metadata, semantic HTML structure, and valid structured data as the three pillars most associated with citation behavior across Brave, Google AIO, and Perplexity. The C-SEO Bench study found traditional SEO-style authority building more effective than content rewriting. The strongest lab evidence for content tactics is the GEO paper (Aggarwal et al., KDD 2024), which reported that adding statistics, quotations and source citations raised a visibility metric by roughly 30-40% on its own GEO-bench benchmark, in a generative engine the authors built, with gains that varied by domain. That is a benchmark result, not a measured change in how often ChatGPT, Gemini or Google cite a brand.

How long does it take for earned media to affect AI citations?

Stacker and Scrunch's March 2026 study of 87 stories, run over 30 days, reported a 239% median citation lift for stories distributed across trusted news outlets. It took one post-distribution snapshot and does not say how soon after distribution the lift appeared. Perplexity updates its source index frequently. The Signal Genesys study, analyzing 179.5 million citation records across 6 LLM platforms, found Perplexity drives the largest citation volume of any AI platform and is among the fastest to index new content from trusted sources.

What counts as a Tier 1 publication for AI citation purposes?

Ahrefs' analysis of the 1,000 pages ChatGPT cited most in September 2025 found that, among the cited pages that also rank in Google, 65.3% sit on domains with Domain Rating 81 or higher (median DR 90). 28% of those 1,000 pages had no organic visibility. Ahrefs measured the domain-rating distribution of pages already in that citation inventory. The study does not establish that Domain Rating causes citation, that Domain Rating 81 is a citation requirement, or that changing Domain Rating changes which sources are selected. Which specific outlets matter varies by industry and query type; the Ahrefs data identifies high-DR domains, not named publications. The measured unit is the domain-rating distribution among pages already present in Ahrefs' citation inventory. Ahrefs does not establish that Domain Rating causes citation, that DR 81+ is required for citation, or that changing Domain Rating changes source selection.

Can B2B brands build earned media authority without a traditional PR agency?

Yes, but the mechanism matters more than the label. Ahrefs' 75,000-brand study measured brand web mentions, not agency retainers. The 0.664 correlation with AI Overview brand visibility reflects the accumulated footprint of brand mentions across the web, whatever generated them. Brands that publish original research, contribute expert commentary to industry publications, or partner with agencies like AuthorityTech for outcomes-based placement can build the earned media signal AI engines weight without traditional retainer PR. The key is that the coverage appears in publications AI engines have learned to trust -- not your own blog, LinkedIn posts, or press release wires.

What is Machine Relations and how does it connect GEO and PR?

Machine Relations is the discipline coined by Jaxon Parrott in 2024 that names the full system of ensuring brands are recognized, cited, and recommended by AI systems. It connects GEO and PR by treating earned media (Layer 1) as the foundation and citation architecture (Layer 3) as the technical extraction layer. GEO operates at Layer 3; PR operates at Layer 1. Machine Relations is the parent framework that makes the dependency explicit: you need both, in the right order, with measurement connecting them.

Correction, November 11, 2026

This page described the Fullintel-UConn analysis as run on a single engine and credited its method to a Fullintel post that gives the source shares but not the method. Fullintel's January 2026 announcement of the analysis gives 400 prompts across 10 personas on weight-loss-drug questions, with the cited articles collected through Scrunch AI, an AI-search monitoring platform; it does not say how many engines. The page now links that announcement for the method. An earlier version of this correction gave instead the method of the study's later stage, which ran a larger prompt set through ChatGPT, Gemini and Google's AI Overviews; that is not the analysis behind these shares. It also gave WorldCom PR Group a count of member agencies that WorldCom's page does not state; that count is removed.