X is where GEO products advertise. It is also where a handful of operators publish measurements that survive a week. We keep the second pile. The first pile is how practices waste a year on llms.txt theater and Reddit threads that ChatGPT stopped displaying.
The overlap is the story
Antoine and others keep circulating the same uncomfortable number: on the order of 11% domain overlap between ChatGPT and Perplexity for the same question. Google’s own AI Overviews and AI Mode are reported around 14%. Alexey Borisov published a same-site mix in August 2026: ChatGPT returned to a handful of canonical URLs; Perplexity spread 12 citations across 13 pages. One template cannot serve both.
- ChatGPT: authority, structure, a few unambiguous answers.
- Perplexity: recency, the answer in the first screen, breadth of distinct URLs.
- Claude: depth and explainers a specialist would not be embarrassed by.
- Gemini: Google entities and YouTube.
- AI Overviews: ordinary SEO plus pages already eligible for a snippet.
Schema is not a magic citation switch
Cyrus Shepard amplified Otterly/Tom Peham data from BrightonSEO: adding markup across 2,000+ URLs correlated with a large lift in Google AI Overview citations and a drop in ChatGPT citations. That is the opposite of a universal GEO hack. Google itself says structured data is not required for generative AI features and there is no special AI schema. We still ship JSON-LD that matches visible copy — for rich results and extractable facts — and we never promise it as a ChatGPT ranking lever.
Do not build on one community
Promptwatch and Qwairy reported ChatGPT’s Reddit citation share falling from roughly 3.8% to 0.5% in August 2026. Retrieval of Reddit pages may still happen; the displayed bibliography changed. Lily Ray has also pointed out that AI Overviews often cite sites talking about an official source while the official source sits in classic results. Third-party mentions matter. A single forum does not.
The commodity-zone test
Apoorv Sharma’s audit is the practical version of Google’s ‘non-commodity’ rule: give Claude only your title, compare the two drafts, highlight everything both contain. That overlap is replaceable. If almost nothing remains, you wrote a page an assistant can generate without you. For YMYL healthcare that is not just an SEO miss. It is a trust miss.
The new moat is being worth mentioning — Niels, August 2026. Packaging one insight into a page, a quotable answer, and a short clip beats three content teams making volume.
What a second pass on X actually added
The first pass was overlap, schema, Reddit, and the commodity-zone test. The second pass is measurement. Tim Soulo’s June synthesis (1B data points) and the August GPT-5.6 rebuild are the two posts that should change a healthcare build.
- 99.9% of AI Overviews are informational. Local ‘near me’ is still Maps. GEO retainers sold against map-pack queries are the wrong product.
- ChatGPT search retrieves via Bing; Gemini via Google; Claude via Brave (Dan Petrovic via Soulo). llms.txt does not inventory your site. Rank in the index the assistant actually searches.
- GPT-5.6 limited itself to trusted domains and dropped displayed Reddit, YouTube, and arXiv share. Lily Ray watched the instant-GEO ads spike in the same week. That is the tell.
- Suganthan: ChatGPT’s new search language has freshness windows and domain targeting. Reddit can be the largest retrieved source and still get zero citations. Vendor pages collect the credits. Astroturf with a 30-day window is designed to fail.
- Aleyda / MJ Cachon: branded prompts fan out ~10 ways, 30% use site:, 24% exact quotes, 10% probe reputation. About pages, NAP, credentials, and clean migrations are the verify step.
- Glen Allsopp / Ahrefs Brand Radar: ChatGPT cited 80 spam sites in a month. Google AI Overviews cited zero. Do not build a clinic’s name on the surface that still cites spam.
- Cyrus / Pew: AI-authored text is everywhere. Empty AI text is the problem. First-party data and a human perspective are not optional on YMYL.
- Ahrefs: YouTube mentions had the highest correlation with AI brand visibility of any factor they studied. Also fragile after GPT-5.6. Cut a useful clip from the same insight. Do not farm it.
Where X conflicts with Search Central, we follow Search Central. Paid ‘get mentioned on the URLs LLMs scrape’ arbitrage is Google’s inauthentic-mentions problem with a new label. Mass ‘best of’ blogs for medical claims are scaled content on a YMYL topic. Blank-page JSON-LD theater is a lie we will not ship. The method page lists the official documents, the studies, and these field notes side by side.