Who AI cites when patients ask about plastic surgery.

When someone considering a facelift or a rhinoplasty asks an AI assistant a question, the answer arrives with sources. We wanted to know whose sources. So we ran the same panel of patient questions through four AI engines on one day, twice, in four metros, and coded every citation in every answer.

The result is not one answer sheet. It is two internets.

The wall

Start with the zeros. ChatGPT answered 102 of these questions and cited YouTube exactly never. Instagram, never. TikTok, never. The same engines’ answers over in Google’s world cited YouTube 71 times and Instagram 73 times. This pattern replicated perfectly across both panels: video and social presence buys visibility in Google’s AI surfaces specifically, and buys nothing at all in ChatGPT.

The wall runs both directions. ChatGPT pays out on a different currency: directories and forums. RealSelf drew 120 citations in ChatGPT, at least one in 45% of its answers, and Reddit another 58. In Google’s engines RealSelf still appears, but it competes with social feeds, Yelp, and Google’s own properties.

Google answers with Google

In AI Mode, the single most-cited domain was google.com itself: 283 citations, mostly links into Google’s internal review viewer showing Business Profile reviews. The right way to read that is not “Google cites itself as a source.” It is that review-intent questions resolve inside Google without linking out. If a patient asks who is reputable, the answer is increasingly assembled from Google’s own review inventory.

Both Google engines also cited individual practice websites heavily, dozens of different clinics’ own domains. ChatGPT almost never did, preferring the directory layer. For a practice, that means your own site can be a citation surface in Google’s AI answers, while in ChatGPT your representation is whatever your RealSelf profile and Reddit presence say it is.

The engine that shrugs

Gemini is the outlier: 107 answers, 210 citations total, and 61 answers, 57 percent, with no citations at all. Scarcity is the finding. There is no reliable citation playbook for an engine that mostly declines to cite.

One observation on what triggers video

The prompts that pulled YouTube citations were not tutorials. They were decision questions: what a facelift costs, how to choose a surgeon at 55, whether anyone has had a bad facelift in Beverly Hills, credential checks. Patient-decision content, not procedure explainers, is what Google’s AI surfaces reach for video to answer.

What this means

The off-site playbook is engine-specific. YouTube and Instagram investment buys Google AI visibility. A maintained RealSelf profile and genuine Reddit presence buy ChatGPT visibility. A blended “be visible in AI” recommendation misdirects spend in both directions, and the size of the wall, 73 citations to zero, suggests the misdirection is not small.

For context: Ahrefs’ June 2026 study of 3M+ US queries found YouTube to be the single most-cited domain in Google AI Overviews overall, at 20.9 percent of mentions. Our vertical snapshot shows the extreme case of that average, and its inversion: in this category the same domain that dominates Google’s AI citations is invisible to ChatGPT.

What this snapshot is and is not

Everything above describes July 23, 2026, and nothing more. One run per panel; AI engines refresh what they cite, and measuring that volatility is the point of running this as a dated series rather than a one-time study. Prompt sets are frozen and publish as-is, flaws footnoted: one prompt carries a stale metro reference from an earlier draft, and the two panels’ prompts are parallel but not word-for-word mirrors, so we make no metro-versus-metro claims. Percentages described as shares of answers count a domain once per answer. Domain classification was done by hand; the classification rules publish with the method note.