Highlights
- The dataset: The analysis behind this study covers over 110,000 AI source citations from prompts specific to the medical practice industry, with AI responses from four AI models run in 12 countries during Q2 2026.
- Reddit makes up 15% of ChatGPT's citations: Its single most-cited source, by a 5x margin over anything else.
- ~81% local-language citation rate in Sweden, ~73% in Germany: But only ~34% in Spain.
- 9% average overlap between any two models: The most divergent pairs share a single domain in their top 20.
- 54% of Microsoft Copilot's citations are from commercial domains: Clinic and directory sites, versus 28% for ChatGPT.
Which sources should you target to get cited as a medical practice brand?
Each model's five most-cited sources, with their share of that model's total citations:
| Model | Top cited sources | Citation character |
|---|---|---|
| Grok | yelp.com (5.5%), rankings.newsweek.com (3.8%), facebook.com (2.6%), reddit.com (2.4%), instagram.com (2.1%) | Global reviews, rankings + social |
| ChatGPT | reddit.com (14.7%), top-rated.online (2.6%), topclinica.com (1.5%), patientguiden.se (1.5%), velja.se (1.3%) | Reddit + niche directories |
| Microsoft Copilot | yably.mx (1.8%), doctoralia.com.mx (1.8%), yably.com.ar (1.7%), vårdcentraler.nu (1.4%), capio.se (1.1%) | Local clinic directories |
| Google AI Overview | instagram.com (7.7%), google.com (4.9%), topdoctors.com.ar (3.5%), facebook.com (2.9%), hospitalitaliano.org.ar (2.3%) | Social + local clinics |
ChatGPT is extraordinarily concentrated, nearly one in seven of everything it cites is Reddit, and no other source breaks 3%. Microsoft Copilot is the opposite: its single most-cited domain is just 1.8%, with citations spread thinly across hundreds of local directories. Grok sits in between, anchored by a handful of global review, ranking, and social platforms. The top cited sources in each list appear in Temso's publisher directory.
How have AI source rankings changed over time in the medical practice industry?
We split the roughly 16-week window at its midpoint (early May 2026) and ranked top domains by their share of citations within each half rather than raw counts. With a single midpoint split, we read the movement as suggestive, not conclusive.
The pattern is stark: global review, ranking, and medical-aggregator platforms that anchored the first half, Yelp's mobile site, US News Health, ScimagoIR, Doctify, Bookimed, YouTube, Germany's Jameda, and Healthgrades, fell out of the top 25 entirely in the second half, while their slots filled with national clinic and doctor directories (unionmd.ca, eldoctor.com.mx, hospitalidee.fr, dcklinieken.nl). Reddit climbed from #5 to the #1 slot. We observe a localizing shift in the citation mix over the window.
What type of content do AI models cite for medical clinics?
We classified every cited domain by category. Commercial content, clinic homepages and for-profit directories, leads overall, which is unsurprising for a local service industry. But the mix differs sharply by model.
Microsoft Copilot is the most commercial model, more than half its citations point to clinic and directory sites, with little UGC or editorial. ChatGPT is the most balanced: commercial, user-generated, reference, and institutional content each take meaningful slices, and it cites far more reference material (14.1%) than Grok or Google AI Overview. Grok is the most editorial-leaning (17.6%, reflecting its appetite for ranking lists and news pages) and the most community-driven alongside ChatGPT. Google AI Overview splits between commercial clinic sites and social UGC. The gap between Microsoft Copilot's and ChatGPT's commercial share, 53.7% vs 28.1%, is large and statistically unambiguous.
Do AI models cite local-language content for medical clinics?
Local sourcing varies sharply by country. Using the source's country-code top-level domain (ccTLD) as a proxy for local-language content, we measured the share of citations pointing to a local-ccTLD source in each non-English market. Spanish-language markets were scored against their own national ccTLD: .ar for Argentina, .mx for Mexico, .es for Spain, and Sweden counts both .se and .nu.
Across all non-English markets, 58.5% of citations point to a local-ccTLD source. But the average hides a sharp split: Germanic, Nordic, and Southern-European markets cluster high (60–81%), while Spanish-speaking markets sit in the mid-30s to mid-40s. The likeliest explanation is structural, not behavioral, Sweden (.se), Germany (.de), and Italy (.it) have dense, mature national directories of doctors and clinics, whereas Spanish-language clinics frequently publish on .com domains. The ccTLD proxy therefore undercounts genuinely local content in markets that favor generic TLDs, so treat the Spanish figures as a floor, not a true localization rate.
Which AI model relies most on local sources for medical clinics?
Holding the market constant, the models differ in how local they go. The heatmap shows each model's local-ccTLD citation rate by language; all reported cells meet the n≈30 sample threshold, and cells too thin to report are left blank.
Microsoft Copilot is the most aggressively local across the board, 80–93% in French, German, Italian, Dutch, and Swedish markets, consistent with its overall reliance on clinic-directory sites. Grok is the most globalized, dipping to 32% in Spanish and 47% in French where it reaches for international review and ranking platforms. ChatGPT lands in a wide band (38–80%), pulled down in Spanish because a large chunk of its citations go to Reddit, a .com that our method counts as non-local everywhere. Google AI Overview's available cells swing widely on smaller bases (87% Italian, 20% Dutch), and its French, German, and Swedish cells are omitted as too thin. The pattern holds: the country drives localization more than the model does, but within a country Microsoft Copilot reaches for local directories while Grok reaches for global platforms.
Should medical practices optimize for each AI model separately?
Almost entirely. We built each model's ranked list of its top-20 most-cited domains and measured how much each pair holds in common, shared domains divided by all the domains the two lists name between them, for all six model pairs. The average overlap is 8.8%, ranging from 5.3% to 17.6%.
No pair shares more than six of twenty domains, and half the pairs share two or fewer. The strongest overlap, Grok and Google AI Overview, rests entirely on global social and review platforms (Facebook, Instagram, YouTube, Yelp's mobile site) plus two Spanish-language directories, not on the local clinic sites that dominate Microsoft Copilot. A medical practice optimizing for one AI model is absent from roughly 91% of the domains the other models prefer.
How many sources does each AI model cite per answer?
The models differ enormously in how many sources they cite per answer. Grok is by far the most source-hungry; ChatGPT the most economical.
Grok cites roughly 5x as many sources per response as ChatGPT. For a clinic, that cuts two ways: Grok offers far more citation slots to compete for, but each individual citation carries proportionally less weight; ChatGPT cites few sources, so earning one of its handful is both harder and more valuable. These averages count only responses that cited at least one source, ChatGPT produced 3,893 responses but cited sources in only 1,093 of them, so most ChatGPT answers cited nothing at all.
Context
This report covers how four AI systems, ChatGPT, Microsoft Copilot, xAI's Grok, and Google AI Overview, cited web sources when answering medical-practice and clinic questions between March 10 and June 29, 2026. The data spans 12 markets (Argentina, Australia, Canada, France, Germany, Italy, Mexico, the Netherlands, Spain, Sweden, the United Kingdom, and the United States) and seven languages, drawn from Temso's AI visibility monitoring platform.
The findings describe which sources AI models cite, not which medical practices or treatments are best. A high citation share means a domain is frequently surfaced by a model, not that it is authoritative or accurate. Citation behavior is also a moving target: models update their retrieval and ranking continuously, so a snapshot like this captures a period, not a permanent state.
Methodology
How we measured this
We tracked four AI models responding to medical-practice and clinic prompts across 12 markets and seven languages over roughly sixteen weeks. Each response was parsed to extract its source citations: the URLs and domains a model explicitly surfaced in support of an answer. We count only sources flagged as cited. Domain categories (commercial, UGC, editorial, reference, institutional) were assigned from a global domain registry, with 8–12% of citations per model left unclassified rather than dropped. Cross-model overlap was measured on each model pair's top-20 most-cited domains, then averaged across all six pairs (mean 8.8%). Temporal analysis split the observation period at its midpoint and compared domain rankings by within-half citation share.
The analysis covers over 110,000 cited source citations from 10,477 AI responses. Localization is measured by a country-code-domain proxy, the source records carry no detected content language, so those figures are directional and exclude English-language prompts, which sit overwhelmingly on generic domains. Because the ccTLD proxy counts .com sites as non-local everywhere, it undercounts local content in markets where clinics favor generic TLDs, so the Spanish-market rates are conservative. A handful of per-model × per-language cells fell below the n≈30 minimum (Google AI Overview's German cell, n=28, and its missing French and Swedish cells) and are omitted from the localization matrix. Model coverage varied across the observation window, so temporal movements are reported as relative ranks within each half; findings are observational and correlational.
Frequently asked questions
Do different AI models cite the same sources for medical-clinic questions?
No, barely. Any two models share only about 8.8% of their top-20 most-cited domains on average, so roughly 91% of the sources one model relies on are absent from another's top tier. There is no single set of "AI-trusted" medical sources; there are several distinct ones.
Which sources does each AI model favor for clinics?
Each model has a distinct profile. Grok favors global review, ranking, and social platforms (Yelp at 5.5%, a Newsweek rankings page, Facebook). ChatGPT is dominated by Reddit, 14.7% of all its citations, plus niche directories. Microsoft Copilot funnels into local clinic and doctor directories (yably.mx, Doctoralia, Capio). Google AI Overview leans on social platforms and Spanish-language clinic directories.
Does local-language content matter for a clinic's AI visibility?
Yes, but the market matters more than the model. In Sweden and Germany, roughly 73–81% of citations go to local-ccTLD sources; in Spain, Mexico, and Argentina, only about 34–45% do. Part of that gap is that Spanish-language clinics often publish on .com domains, which our method counts as non-local, so those figures are conservative.
Which AI model is most likely to cite a local clinic's website?
Microsoft Copilot. More than half its citations (53.7%) are commercial clinic and directory sites, and it reaches 80–93% local-ccTLD rates in French, German, Italian, Dutch, and Swedish markets. If your priority is being cited as a local practice, Microsoft Copilot is the most receptive system.
Which AI model cites the most sources, and does that help my clinic?
Grok, by a wide margin, about 29 sources per response, roughly five times ChatGPT's 5.4. More sources mean more slots to compete for in Grok, but each carries less weight. ChatGPT's economy of citations, and the fact that most of its answers cite nothing at all, makes each of its few slots both harder to win and more valuable.
Is being on Reddit worth it for a medical practice?
For ChatGPT, enormously. Reddit is ChatGPT's single most-cited domain at 14.7% of all its citations, more than five times its next source, and it ranks among Grok's top sources too. User-generated discussion is a real path to AI visibility, but mainly for the models that lean on it; Microsoft Copilot and Google AI Overview barely cite it.
Did the sources AI models cite change over the period?
The citation mix shifted toward local clinic and doctor directories in the second half, as global aggregators like Yelp's mobile site, US News Health, Doctify, Bookimed, and Healthgrades dropped out of the top tier. We read it as a suggestive shift, not a proven trend.
What should a medical practice do with this?
Optimize per model, not in general. There is no universal AI-visibility strategy here: presence on local national directories helps with Microsoft Copilot, global review and ranking platforms help with Grok, social profiles plus local directories help with Google AI Overview, and a strong Reddit footprint plus reference content helps with ChatGPT. Treat each model as a separate channel with its own preferred sources.
