84% of What AI Engines Cite Isn't From Your Website

The standard AI visibility checklist starts with your website: fix your schema, update your service pages, add a location page. Run the audit, find the gaps, patch them.

That checklist is not wrong. But it addresses the minority of what AI engines actually look at. Two independent studies published in May 2026 -- one analyzing over one million AI prompts -- found that 84 to 85.5 percent of what AI engines cite comes from third-party sources, not the brand's own website.

That figure reframes the entire optimization problem.

The earned media finding

In our June 2026 investigation of content signals for AI platforms (session 43), we documented two converging datasets.

Muck Rack analyzed over one million AI prompts and found that 84% of AI citations reference earned media sources -- third-party coverage, directory listings, review platforms -- rather than brand-owned content. This is a direct measurement of citation source type, not a proxy metric.

5W PR published a separate earned media study on May 15, 2026, with a near-identical finding: 85.5% of AI citations come from earned media. The same study cited University of Toronto research showing AI engines cite earned media five times more frequently than brand-owned websites. The multi-publication effect was more striking still: spreading content across multiple third-party publications increases AI citations by 325% compared to publishing the same content on the brand site alone.

The mechanism is not hard to understand once you see it. AI platforms have spent years learning not to trust self-reported brand claims. A page that says "we are the best plumber in Toronto" carries no external signal. A Yelp profile with 47 verified reviews, a Clutch entry with project history, and a mention in three local roundup articles all carry signal that the brand website cannot replicate. The platforms are systematically weighting sources that have been corroborated by third parties -- not sources that assert their own authority.

Which third-party sources actually get cited

The 84% finding obscures something important: "third-party sources" means different things on different platforms, and the mix is not uniform.

Our May 2026 platform citation behaviors investigation (session 14) documented the Yext 2026 AI Visibility study -- an analysis of 17.2 million citations across ChatGPT, Perplexity, and Gemini. The source architecture splits cleanly by platform.

ChatGPT pulls 49% of its citations from third-party directories: Yelp, BBB, G2, Capterra, industry registries. ChatGPT relies heavily on training data rather than live retrieval, so these directories must be indexed before training cutoffs. Appearing in them now affects future training cycles; existing directory presence is already working or not.

Perplexity uses niche expert directories specific to each industry. For lawyers, that means Avvo and Martindale-Hubbell. For home services, Thumbtack and Angi. For SaaS, G2 and Capterra. The general-purpose directories that satisfy ChatGPT don't fully transfer to Perplexity. Perplexity also retrieves live web pages, so crawlability and content freshness matter more here than on ChatGPT.

Gemini is the platform where brand-owned content matters most: 52% of its citations come from brand-owned websites, making it the exception to the earned media dominance pattern. Strong on-site structure -- schema markup, E-E-A-T signals, a well-indexed Google Business Profile -- matters more for Gemini than for the other two.

One source cuts across all of them. BrightLocal's 2025 analysis, documented in our April 2026 directory presence investigation, found that Yelp appears in 33% of all AI search citations across platforms, and that Perplexity uses Yelp as a source in every single industry category they tested. There is no other directory that appears across that range. Yelp is the one directory that doesn't require a platform-specific strategy.

The publication reranking story (for businesses where press is attainable)

For businesses where press coverage is achievable -- B2B, professional services, SaaS, healthcare -- there is a second layer to the earned media question: which publications AI engines actually cite.

Our June 2026 investigation (session 48) found the 5W Trade Press AI Index 2026, a cross-industry study of 680 million citations across ChatGPT, Claude, Perplexity, Gemini, and Google AI Overviews. The finding: AI engines have quietly reranked the trade press in ways that don't track readership, prestige, or advertising revenue.

For technology, PCMag leads AI citation share. TechCrunch is explicitly documented as losing citation ground despite its audience size and editorial reputation. For healthcare, STAT outperforms general health publications. For financial services, Bloomberg leads over broad business outlets.

The pattern is consistent with other citation evidence we've tracked: AI engines favor precision, factual density, neutral voice, and topical specificity. PCMag's structured product reviews, STAT's citation-heavy medical reporting, and Bloomberg's sourced financial coverage all match what AI retrieval layers reward. TechCrunch's narrative tech journalism does not, even though it reaches more readers.

For most SMBs, trade press coverage is not on the table. But for any business where it is -- a legal practice, a healthcare provider, a regional accounting firm, a B2B vendor -- "get press coverage" is now too vague to be useful. The outlet selection matters as much as the coverage itself.

What a minimum viable third-party footprint looks like

For local and small businesses, the earned media imperative translates to a specific set of assets.

Our April 2026 directory presence investigation found that every client who scored zero on category authority queries had minimal or no third-party citation footprint. The competitors outperforming them on Perplexity -- showing up when someone asks "best [service] in [city]" -- consistently had established footprints across review platforms and niche directories.

The minimum viable footprint, based on what we observe across audits:

Yelp is the non-negotiable first step. It appears in 33% of all AI citations and in every Perplexity industry category. No other directory is this universal.

Google Business Profile anchors Gemini performance. GBP data feeds directly into Gemini's local responses. For Gemini specifically, this is the single most important directory asset.

One or two industry-specific directories with actual reviews -- not claimed-but-empty profiles. For healthcare, that means Healthgrades or Zocdoc. For home services, Angi or Thumbtack. For agencies and professional services, Clutch. For Perplexity specifically, the review content on Yelp and niche directories appears to be citation material, not just listing metadata. An empty profile gives Perplexity nothing to synthesize from.

LinkedIn rounds it out. It is the third most-cited domain across AI platforms in Peec's 30-million-source study, and it applies to nearly every business category.

That is roughly five independent, consistent source references. The research notes that "businesses with 10 accurate, high-authority citations consistently outrank competitors with 200 low-quality inconsistent ones" -- so breadth without quality doesn't compound the way quality concentration does.

What this doesn't mean for your website

The 84% figure does not make on-site work irrelevant. Schema markup, service page structure, and Google Business Profile optimization all still move scores -- Gemini especially relies on brand-site signals, and the on-site work matters for crawlability across all platforms.

What it does mean: the optimization problem is primarily an off-site problem. Most of what AI engines need in order to cite a business is not on the business's own domain. A fix plan that starts with schema and stops there is addressing the 15% before touching the 85%.

The Signal Check audit shows where your third-party footprint is missing, which platforms are failing because of it, and what the priority sequence looks like starting from the actual source of the gap -- not from the easiest things to change.