GEO courses quote Ahrefs the way they used to quote a Moz study: one chart, a product, a waitlist. Tim Soulo’s June 2026 thread is the opposite. Fourteen public studies. A billion data points. Most of it is bad news for people selling a special AI file. It is good news for a practice that already needed a serious website.
The numbers we will stand on
- Adding JSON-LD on 1,885 pages, matched to 4,000 controls: ChatGPT +2.2%, AI Mode +2.4%, AI Overviews −4.6%. That is noise. Schema is not a citation switch.
- YouTube mentions correlated 0.737 with AI brand visibility — stronger than backlinks, page count, or domain rating — on both Google-owned products and OpenAI.
- 67% of ChatGPT’s top 1,000 citations are Wikipedia, homepages, or app stores. Marketers do not ‘optimize’ those. 32% are influenceable: educational pages, reviews, news, blogs.
- 28.3% of ChatGPT’s most-cited pages have zero Google organic visibility. ChatGPT is not a reranker of your blue links.
- ChatGPT cites about 50% of the URLs it retrieves. Being fetched is not being named.
- AI Overviews cut clicks to the #1 result by 58%, up from 34.5% ten months earlier.
- 99.9% of AI Overviews fire on informational intent. Local, navigational, and transactional searches are almost AIO-free. Shopping: 3.2%.
- AI Mode and AI Overviews reach the same conclusions 86% of the time and share 13.7% of citations.
- An Overview reshuffles wording every ~2.15 days (70% of tokens differ) while semantic similarity stays at 0.95. Chase the meaning.
What this means for a medical practice
Two different jobs got sold as one ‘AI SEO’ retainer. Job one is informational: ‘how long does Invisalign take,’ ‘is PRP safe after 50,’ ‘what does a pelvic-floor PT actually do.’ That is Overviews, ChatGPT, Perplexity, Claude. Job two is local and transactional: ‘Invisalign dentist Minneapolis,’ ‘book a same-day crown,’ ‘PT near me.’ Ahrefs says that job almost never triggers an Overview. Maps, reviews, NAP, a phone that works, and a page that matches the query still book the visit.
If an agency is selling you GEO for near-me queries, they are selling the wrong surface. If they are not building condition and service explainers for the informational side, they are leaving the assistants empty-handed.
Schema, again, because someone will invoice it
Otterly / BrightonSEO data that Cyrus Shepard amplified showed markup correlating with an Overview lift and a ChatGPT drop. Ahrefs’ later controlled test showed nothing. Google’s own 2026 guide says structured data is not required for generative AI features and there is no special AI schema. Our rule does not depend on which dataset you prefer: JSON-LD that matches visible copy, so a model can extract specialty, city, hours, and providers. Never billed as a ChatGPT ranking lever.
YouTube is not a side hustle
The 0.737 correlation is the strongest Ahrefs published in that set. Gemini especially. It is also fragile: GPT-5.6 cut YouTube’s displayed citation share in the same week it cut Reddit and arXiv. So we do not farm faceless listicles for a citation. We cut a short native clip from the same insight as the service page — the clinician, the method, the room. Packaging, not a second content team.
The 67% you cannot buy
Wikipedia, the homepage, the app store. That is where ChatGPT’s head citations live. For a clinic: a homepage that states who you are in the first screen, a Google Business Profile that is complete, and third-party pages a model already trusts (Healthgrades, a specialist society, a real newspaper, a Wikipedia page you did not ghostwrite). The remaining 32% is the work: educational service pages, reviews, news. ‘Best dentist 2026’ farms are how YMYL sites get into spam trouble. Lily Ray has been saying the listicle trick is already fading in GPT-5.6. We believed her before the chart.
Retrieved is not cited. ChatGPT fetched 84 Reddit threads for one answer and named none of them. The vendor pages collected the credits. — Suganthan Mohanadasan, August 2026, on his own traffic.
How we measure, since clicks will keep falling
Soulo also noted AI-referral traffic to 75,000 sites fell 3% over eleven months while AI adoption rose. Chatbots are walled gardens. A mention can live on a third-party URL. So we sample the questions a practice actually gets asked, by engine, every month. We still watch Search Console’s generative-AI report, GBP calls, and tel: taps. Revenue is the metric. Share-of-answer is a leading indicator. A GEO dashboard with no booked visits is a screensaver.
The method page lists the official Google documents, these studies, and the X field notes side by side. Where they conflict, Search Central wins. Where Search Central is silent, we use the study. Where the study is a course funnel, we ignore it.