Earned Media Beats Owned Content in AI Search: What a 2025 Generative Engine Optimization Study Found

What the 2025 study actually found

In 2025 a research team led by Chen and colleagues published Generative Engine Optimization: How to Dominate AI Search (arXiv:2509.08919), a large-scale controlled follow-up to the earlier academic work that coined the term. Its central reported result is a source-selection bias: across the systems it examined, generative answer engines drew disproportionately on earned and third-party media, the independent publications, industry resources, and outside authorities that talk about a business, rather than on the brand-owned pages a business publishes about itself or the social content it posts.

What makes the finding notable is the comparison it draws. The same study contrasts that lopsided behavior with classic Google search, which the authors report sources more evenly across owned and earned material. In other words, the tilt toward third-party authority is not simply how all search has always worked; it is reported to be sharper in the generative-answer layer than in the ranked-link layer that preceded it.

The authors tested the pattern across multiple verticals, multiple languages, and paraphrased versions of the same underlying query, which is a meaningful design choice: it guards against the finding being an artifact of one industry, one phrasing, or one language. That breadth is why the result is worth taking seriously even while it remains a single study.

Earned media and owned content are two different levers

The distinction at the heart of this piece is old in communications, even if it is newly consequential in search. Owned content is everything a business controls and publishes itself: its website, its landing pages, its blog, its own social accounts. Earned media is what independent third parties say about it without being paid to say it: a journalist's article, an industry directory listing, a review corpus, an expert citation, a reference in a resource other people maintain.

The two levers behave differently because their credibility comes from different places. Owned content is a first-party claim; its author and its subject are the same party. Earned media is a second-party endorsement; the source has no obligation to be flattering, which is precisely what gives it evidential weight. A system trying to answer "who is good at this" has a structural reason to lean on the second kind of signal, because it is harder to manufacture.

This is not the same as the paid-media lane. Buying links, renting placements, or manufacturing coverage is neither earned nor durable; it violates search platforms' link-spam policies and is increasingly filtered out of AI citations. The lever the evidence points to is genuine earned authority, mention won on merit, not purchased corroboration dressed up to look like it.

The founding GEO experiment pointed the same way

The 2025 result does not sit in isolation. The peer-reviewed study that established generative engine optimization as a field, Aggarwal and colleagues at ACM SIGKDD 2024, ran a controlled benchmark of roughly 10,000 queries across nine datasets and measured which content-level changes moved whether a source was surfaced inside a generated answer.

Its strongest levers are directly relevant here. Adding citations to credible sources, including direct quotations, and replacing vague claims with concrete statistics produced a relative lift on the order of 30 to 40 percent on the study's visibility metric versus unoptimized baselines, and citing authoritative sources was consistently the single most effective intervention it tested. That earlier experiment measured the tactic (borrow authority through citation) while the 2025 study measures the pattern (engines prefer authority they can borrow). They are two readings of the same underlying pull toward corroborated, attributable material.

When an independent 2024 result and a large-scale 2025 result point in the same direction from different angles, the convergence is worth more than either finding alone, even though neither, on its own, closes the question.

Why retrieval would reward corroboration over self-assertion

A plausible mechanism sits underneath both findings, and it is worth stating carefully as reasoned inference rather than proven fact. Nearly every generative answer engine is built on retrieval-augmented generation: the founding architecture, described by Lewis and colleagues in 2020, pairs a language model with a searchable index of outside documents that is queried at answer time, which the authors showed produces more specific and factual output than a model relying on its parameters alone.

In that design, the answer is assembled from whatever passages the retrieval step surfaces as most relevant and trustworthy for the query. A page that asserts its own excellence offers the retriever a first-party claim. A cluster of independent pages that corroborate the same fact offers something a ranking or relevance system has long treated as stronger: agreement from sources that do not share the subject's incentive. It is reasonable to expect a retrieval-first system to over-index on that corroboration, and the two studies above are consistent with exactly that expectation. The mechanism is inferred from the architecture and the results; it has not been isolated experimentally, and we flag it as such.

How this departs from classic search

For two decades, a business could win a great deal of search visibility on the strength of its own site. Publish enough well-structured, useful pages, earn a reasonable link profile, and you could rank. Owned content was the primary lever, and earned media was an accelerant. The 2025 study's reported contrast, that classic Google sources more evenly while generative engines tilt harder toward third parties, suggests that habit transfers poorly to the answer layer.

The consequence is concrete. A business can hold a strong classic ranking and still be missing from the AI answer written above or instead of those links, because the two surfaces are drawing on different mixes of sources. This is one more piece of evidence that classic search visibility and AI-answer visibility are decoupling into related but separate disciplines, which is why measuring them as one blended number obscures more than it reveals.

The limits of this finding

Rigor here means naming the limits as plainly as the finding. First, the 2025 study is a single, not-yet-independently-replicated result. Its breadth across verticals and languages strengthens it, but one paper, however large, is an emerging signal, not an established law. Anyone selling the earned-media tilt as a settled certainty is ahead of the evidence.

Second, being cited is not the same as being read faithfully, nor the same as being visited. The RAG faithfulness literature shows that in the dominant production pipelines, a citation can be attached to an answer independently of the passages the model actually reasoned from; one 2024 attribution survey reported that more than 95 percent of answers from the open-source systems it tested contained at least one unattributed sentence. A citation is a claim of provenance, not a guarantee of it.

Third, even a citation that is faithful may never turn into a click. Pew Research Center's 2025 behavioral study of 68,879 real Google searches found users clicked a traditional result in about 8 percent of searches where an AI summary appeared, versus about 15 percent without one, and clicked a link inside the summary itself only around 1 percent of the time. Winning earned citations is necessary for AI-answer presence; it is not, by itself, a traffic strategy. The right yardstick for this surface is presence in the answer, measured directly, not organic sessions.

What the study does not claim

It is as important to state the boundaries of the result as its substance. The study does not claim that owned content no longer matters; a business still needs a coherent, well-structured, machine-readable site for any engine to have a first-party entity to attach earned mentions to. It does not provide a public per-vertical citation rate that a specific business can benchmark itself against. It does not prove that buying coverage works; the durable lever is earned, not purchased, and manufactured corroboration is both against platform policy and increasingly filtered out.

And it does not promise that a given amount of earned media produces a given amount of AI visibility. There is no guaranteed dose-response curve in the evidence. What the study supports is a directional reweighting of strategy, not a formula.

What this means for a visibility strategy

Read together, the evidence argues for a shift in where effort goes. A visibility program built almost entirely on owned pages is optimized for the surface that is losing share of the decision. Reallocating some of that effort toward earning genuine third-party authority, the digital-PR discipline of building citable assets and placing them on sources engines already trust, aligns the strategy with how the answer layer appears to select what it cites.

The reallocation is a rebalance, not a reversal. Owned foundations still carry classic search and give earned mentions an entity to point at; earned authority is what the generative layer disproportionately reaches for. The pairing is the point. And because none of this can be assumed, the discipline that ties it together is measurement: sampling real buyer questions across each engine and recording how often a business is actually named, then watching whether earned placements move that number over time, reported with variance rather than asserted.