Here is the most quietly brutal statistic in AI search: according to Muck Rack's analysis of more than 25 million links cited by ChatGPT, Claude, and Gemini across 17 industries, earned media — independent editorial coverage — accounts for 84% of AI citations. Brand-owned websites, the asset most companies spend the most money on, supply roughly 5–10% (McKinsey, August 2025, n=1,927). Paid and sponsored content, the asset companies spend the second most money on, accounts for 0.3%.
Read those three numbers again as a budget allocation review. In the citation economy, the content you control least is worth roughly 280× the content you pay to place, and about 10× the content you publish yourself. This article explains why the economics work this way, where the exceptions are, and what a rational strategy looks like once you accept the math.
The Market Shares of Attention
Three datasets, collected independently, converge on the same source hierarchy:
| Source type | Share of AI citations | Study |
| Earned media (independent editorial, expert commentary) | 84% — stable between 82% and 89% across three research waves since July 2025 | Muck Rack Generative Pulse, 25M+ cited links, 17 industries (May 2026) |
| Brand-owned websites | 5–10% | McKinsey survey analysis, August 2025 |
| Paid / sponsored content | 0.3% | Muck Rack Generative Pulse |
Within the earned-media bucket, news coverage alone supplies 27% of AI citations. And the academic work agrees with the industry data: Chen et al., How to Dominate AI Search (arXiv:2509.08919), ran controlled experiments across verticals and languages and found AI search systems over-index earned media relative to brand-owned and social content far more aggressively than Google's classic results, which maintain a more balanced source distribution.
This is not a rounding error or an artifact of one vendor's methodology. It is the shape of the new market.
Why Engines Prefer Strangers to Owners
The bias is structural, not accidental. Three mechanisms produce it.
1. The verification architecture of generative systems. An AI answer that cites only the brand being asked about is, by construction, an advertisement. Systems trained to be helpful and hedge against manipulation do better when independent sources corroborate a claim. A third party saying "X is the leading option for Y" carries evidentiary weight that the same sentence on x.com does not. The model's training and the retrieval layer both reward independent corroboration — which is precisely what "earned" means.
2. The fingerprints of the training corpus. Each engine's citation behavior reflects what it reads most. ChatGPT's top-10 cited sources are nearly half Wikipedia (47.9%); Perplexity cites Reddit in 46.7% of relevant-topic answers; AI Overviews lean on Reddit (21%), YouTube (18.8%), Quora (14.3%), and LinkedIn (13%) (Amsive, 2025–2026). Reddit is the single most-cited domain across five major AI platforms per Peec AI's 30-million-citation analysis (March 2026). None of those are places where your brand's marketing team publishes the content.
3. Big brand bias — the second filter. Chen et al. document that AI search systematically favors large, established brands over smaller players — a bias smaller brands must actively work against, not merely acknowledge. Amsive's category analyses make the concentration visible: Amazon holds 57.3% AI visibility in its category, Bank of America 32.2%, Mayo Clinic 14.1%, Harvard 20.8%. Earned media dominance and big-brand dominance compound: the brands journalists already cover are the brands AI engines already trust.
The Exception That Should Organize Your Calendar
Before this reading of the data turns fatalistic, one important nuance from BrightEdge's mid-funnel research (eight industries, AI Overviews and ChatGPT): at the consideration stage of the buyer journey — where AI search demand runs 4%–26% of queries depending on industry — brand-owned content captures 42%–79% of citations. Third-party review aggregators, by contrast, get 1%–7%.
The pattern is intuitive once you see it. Top-of-funnel questions ("what is X," "how does X work") get answered from encyclopedias, forums, and press. But when a user asks an AI engine "how does [brand]'s platform handle Y" or "pricing for [brand] vs [competitor]," the engine has no choice but to consult the brand's own buying guides, documentation, and comparison pages — because that is where the specifics live.
The strategic implication: your owned content matters most exactly where the money changes hands. If your site's AI-facing content portfolio is all top-of-funnel thought leadership, you are competing in the 5–10% bucket with material built for a layer you don't win. If it includes deep comparison pages, buying guides, documentation, and pricing transparency, you are playing in the one layer where owned content is the majority citation source.
Mention Engineering: The New Link Building
If earned mentions are the currency, the rational response is to treat mention acquisition with the discipline SEO teams once reserved for links. The evidence base for that discipline is already substantial:
The correlation gap is wide. Ahrefs' 75,000-brand study (August 2025) found unlinked brand mentions correlate with AI Overview citations at a Spearman coefficient of 0.664, versus 0.218 for total backlinks — roughly 3×. YouTube brand mentions showed the strongest single signal (~0.737), Reddit mentions ~0.674, Wikipedia mentions ~0.659. Top-quartile mention volume brands averaged 169 AI Overview mentions; the next band down averaged 14. These are correlations, and Ahrefs itself flags them as such — but the direction and magnitude are consistent across every independent dataset in this article.
The formats that earn citations are known. Backlinko's June 2026 LLM Seeding analysis identifies seven structures AI engines disproportionately cite: structured "best of" lists with documented methodology; first-person product reviews with reproducible tests; brand comparison tables with per-use-case verdicts; FAQ-format content with question-headed sections; opinion pieces with author credentials and evidence; visual content with full alt-text; and free tools and templates. Listicles alone account for 32% of all AI citations (SEOMator, 177 million citations analyzed) — a share no other single format approaches.
The platforms that generate citations are known. Reddit (the #1 cited domain across platforms), Quora (the most-cited source in AI Overviews), G2 and Capterra (where comparison-stage buyers verify), YouTube (whose transcripts feed both Google's and OpenAI's pipelines), LinkedIn, and expert commentary in trade press. Growth Memo's August 2026 research adds a finding worth underlining: UGC platforms earn citations at every stage of the buyer journey at higher rates than dedicated review sites.
And the queries where brand mentions occur are known. Moz's 50,000-fan-out study (August 2026) found entity prompts ("what is X") and comparison prompts ("X vs Y") account for 97% of all brand mentions in generative outputs. Absent from both? You are absent from the answer — full stop.
Assemble those four facts and mention engineering stops being a buzzword: it is a content calendar (seven formats), a channel plan (six platform classes), and a query map (entity + comparison), executed with PR discipline and measured in rolling windows.
What This Doesn't Mean
Three honest limits before anyone reallocates the entire budget.
Correlation is not causation. Brands with heavy earned mentions are frequently brands with heavy everything — search demand, direct traffic, incumbency. The Ahrefs coefficients describe an ecosystem, not a lever you can pull once and watch the dial move.
Earned media is slow, and that is its price of admission. Unlike paid distribution, press coverage and Wikipedia presence cannot be bought directly, and attempts to game them (fabricated reviews, astroturfed Reddit threads) carry a different kind of risk: the platforms involved police manipulation actively, and a brand caught manufacturing "earned" signals loses the trust that earned media exists to convey.
Small brands start with a handicap, not a wall. Chen et al.'s big-brand bias is real, but the same paper frames it as something niche players offset with more aggressive mention strategies — not an impossibility result. The Amsive data showing anomalously high LLM visibility for smaller brands in some financial categories (Navy Federal, Upstart) suggests the AI layer is less locked-down than the classic SERP, where a decade of accumulated domain authority decides most contests.
The Honest Bottom Line
The 84% number is not an argument for publishing less on your own domain. It is an argument for recognizing what each asset class is actually for in the citation economy: owned content wins the consideration layer, where specificity beats independence; earned media wins everything above it, where independence is the value; paid content barely exists at all.
Most marketing organizations have the allocation inverted — heavy on owned and paid, light on earned — because earned is the hardest to buy and the slowest to compound. The data says it is also the only layer with a structural moat. The brands that internalize this first will spend the coming few years collecting mentions the way their competitors spent 2015 collecting links: awkwardly, gradually, and then all at once.
Sources
Muck Rack (2026). Generative Pulse. 25M+ cited links, 17 industries; three research waves, July 2025 – May 2026.
McKinsey (2025). AI search citation survey analysis. n = 1,927, August 2025.
Chen, X., et al. (2025). "Generative Engine Optimization: How to Dominate AI Search." arXiv:2509.08919.
BrightEdge (2026). Mid-funnel AI citation research. Eight industries; AI Overviews and ChatGPT.
Ahrefs (2025). Brand mentions and AI visibility. 75,000 brands; Spearman correlations.
Backlinko (2026). "LLM Seeding: A New SEO Strategy to Get Mentioned by LLMs."
Amsive / Lily Ray (2026). AI citation source-structure analyses; category visibility tracking.
Peec AI (2026). 30 million citation analysis. March 2026.
SEOMator (2026). 177 million citation format analysis.
Moz / Dr. Peter J. Meyers (2026). "What 50k Query Fan-Outs Reveal About Brands."
Growth Memo / Kevin Indig (2026). Community signals research; UGC vs. review-site citation rates. August 2026.
All correlation figures are reported as correlations. Figures reflect research published as of September 2026; source shares shift as engine retrieval stacks evolve.