AI Cites Lists, Not Homepages, and ChatGPT and Claude Don't Read the Same Web
We mapped 63,500 AI citations across 10,000 commercial prompts in 50 industries. The pages that win are ranked best-of lists, and ChatGPT and Claude pull from very different corners of the web.
TL;DR: We mapped 63,500 AI citations across 10,000 commercial prompts in 50 industries with our research partner citations.press. The pages that win are ranked best-of lists (about three in four citations), homepages barely register, and ChatGPT and Claude pull from very different corners of the web.
Every brand chasing AI visibility asks the same question: what do we publish to get cited? Most answer with instinct, more blog posts, more homepage copy. Our analysis at LLM Scout says almost all of that misses.
We examined 63,500 citations that ChatGPT and Claude would produce across 10,000 real commercial prompts in 50 industries, from CRM and cybersecurity to insurance, healthcare, electric vehicles and coffee. Two patterns are strong enough to change how you brief content: the format that wins is the ranked list, not the homepage, and ChatGPT and Claude do not cite the same web.
Dataset and Analytical Scope
This is based on structured citation modelling, not a survey. The study covers 10,000 commercial queries across ten intent types (best, compare, guide, selection, reviews, trends, features, buying, recommendations, evaluation), 50 industries with 20 prompts each, and ChatGPT plus Claude for 63,500 citations examined. Metrics include page type cited, domain concentration, cross-model overlap, and source archetype by model.
Note. Citations are model-estimated, not a scrape of live answers. A Gemini pass was excluded because it mirrored ChatGPT and carried no independent signal. Page types were inferred from URL structure, so bucket sizes are directional.
The Format That Wins Is a List, Not a Homepage
Group every cited URL by the kind of page it is and one format dominates both models. Ranked best-of and comparison pages are the single largest category, and adding the category and directory listing pages that behave the same way, list-style pages account for 81% of ChatGPT citations and 75% of Claude citations.
| Page type | ChatGPT % | Claude % |
|---|---|---|
| Best-of / comparison | 36 | 58 |
| Category index | 35 | 9 |
| Standard article | 12 | 21 |
| Directory / listing hub | 11 | 8 |
| Homepage | 7 | 4 |
The takeaway is simple: AI reaches for pages structured as ranked comparisons of named options, because that structure maps onto how a buyer phrases the question. "Best CRM for small business" has an obvious answer shape, and a ranked list fills it. A homepage does not.
Homepages Barely Register
This is the finding most brands resist, because most AI content effort still points at the homepage. Here the homepage is the least cited page type, 7% for ChatGPT and 4% for Claude.
The strategic implication: the citation you want is usually not on your own domain. It is a slot on a publisher's ranked list.
Example Prompts and What Each Model Cited
To make the divergence concrete, here are ten prompts from the dataset and the three domains each model would cite. Some questions produce near-identical answers; others send the two models to completely different sources.
| Prompt | ChatGPT cited | Claude cited |
|---|---|---|
| Best CRM software | techradar, g2, gartner | forbes, zapier, techradar |
| Best cybersecurity | capterra, softwareadvice, g2 | techradar, cybernews, security.org |
| Best healthcare options | nhs.uk, who.int, cdc.gov | forbes, nerdwallet, insurance.com |
| Best electric vehicles | caranddriver, whatcar, edmunds | edmunds, kbb, cars.usnews |
| Best mortgages | moneysavingexpert, which.co.uk, moneyhelper | nerdwallet, forbes, bankrate |
| Best universities | topuniversities, timeshighered, usnews | usnews, timeshighered, topuniversities |
| Best VPN | techradar, tomsguide, pcmag | techradar, tomsguide, cybernews |
| Best coffee | sca.coffee, ncausa.org, seriouseats | consumerreports, tasteofhome, themanual |
| Best gaming | ign, gamespot, pcgamer | ign, gamesradar, nintendolife |
| Best insurance in the UK | nerdwallet, moneysavingexpert, which.co.uk | nerdwallet, forbes, bankrate |
On "best universities" and "best VPN" the models converge. On "best healthcare" they split entirely: ChatGPT answers with official bodies (NHS, WHO, CDC) while Claude answers with commercial comparison sites. On "best coffee" there is zero domain overlap.
How Concentrated Is the AI Web?
Both models funnel citations into a small set of domains, but ChatGPT funnels harder. Its top ten domains capture 35% of all its citations, and just 20 domains cover half of everything it cites. Claude is flatter: its top ten take 26%, and it needs 46 domains to cover half.
| Concentration metric | ChatGPT | Claude |
|---|---|---|
| Unique domains cited | 110 | 208 |
| Top-10 domains' share of citations | 35% | 26% |
| Domains needed to cover 50% of citations | 20 | 46 |
In practical terms, a smaller club of publishers controls a larger share of ChatGPT's answers, while Claude distributes attention across a longer tail.
How Often the Two Models Agree
Across the 10,000 prompts, ChatGPT and Claude produced the same three-URL answer exactly zero times. Measured by shared domains, they overlapped on just one source or fewer on 90% of prompts, and agreed on all three sources on under 1%.
| Shared domains (of 3) | % of prompts | Approx. prompts | Read |
|---|---|---|---|
| 0 (no overlap) | 55% | 5,500 | No shared domains on that prompt |
| 1 shared | 36% | 3,600 | Exactly one domain in common |
| 2 shared | 9% | 900 | Two of the three sources overlap |
| 3 (identical) | 1% | 100 | All three sources match |
The Pattern Holds Across Every Question Type
The list-page dominance is not confined to one kind of query. Across all ten intent types we tested, ChatGPT's mix stays remarkably stable: best-of and category pages together sit around 70% on every intent, and homepages never rise above 7%.
| Prompt intent | Best-of % | Category % | Homepage % |
|---|---|---|---|
| Best | 37 | 32 | 6 |
| Compare | 35 | 37 | 7 |
| Guide | 37 | 32 | 6 |
| Selection | 36 | 35 | 7 |
| Reviews | 36 | 34 | 6 |
| Trends | 35 | 37 | 7 |
| Features | 37 | 33 | 6 |
| Buying | 35 | 37 | 7 |
| Recommendations | 36 | 34 | 6 |
| Evaluation | 36 | 35 | 7 |
Where Official Sources Take Over
The averages hide sharp swings by industry. In healthcare, 100% of ChatGPT's citations came from official or authority sources such as the NHS, WHO and CDC. Dentistry and nutrition are close behind. At the other end, some categories are dominated by a handful of destination sites cited at their homepage.
| Signal (ChatGPT) | Industries most affected |
|---|---|
| Highest authority-source use | Healthcare (100%), Dentistry (78%), Nutrition (75%), Legal (52%), Parenting (52%) |
| Highest homepage use | News (75%), Logistics (51%), Airlines (27%), Beauty (27%), Restaurants (27%) |
| Highest UK-source use | Payroll (100%), Ticket Resale (75%), Legal (75%), Real Estate (75%), Investing (51%) |
The Same Question, Two Different Webs
Format is where the models agree. Sourcing is where they split. Claude drew on 208 unique domains to ChatGPT's 110, so Claude is broader and longer-tail while ChatGPT concentrates on a few mega-hubs.
ChatGPT leans on TechRadar, PCMag and G2 category pages. Claude leads with Forbes Advisor and TechRadar but pulls a wider editorial and analyst set, including Zapier, Cybernews, Deloitte and Consumer Reports. A shared spine (TechRadar, Forbes, G2, Which?, NerdWallet) is trusted by both, but beyond it the two diverge quickly.
Citation Personality: Authority vs Editorial
| Signal | ChatGPT | Claude | Read |
|---|---|---|---|
| Authority (gov, health, reg) | 14% | 8% | ChatGPT de-risks toward official sources |
| UK sources | 18% | 10% | ChatGPT surfaces UK sites about 70% more |
| B2B directories | 11% | 8% | Both use them; ChatGPT slightly more |
| Unique domains | 110 | 208 | Claude spreads far wider |
Put plainly: ChatGPT behaves like a cautious librarian, favouring official, high-trust and regional sources. Claude behaves like a well-read analyst, favouring a broad editorial and consultancy base.
Two Playbooks, Not One
| If your audience uses... | Prioritise | Because |
|---|---|---|
| ChatGPT | Ranked lists on TechRadar/PCMag/G2; official or UK corroboration | It concentrates on a few authority hubs and trusts official/regional sources |
| Claude | Broad editorial coverage; analyst and research mentions for trends | It spreads across 200+ domains and leans on editorial and consultancy content |
| Both | Featured on the shared spine: TechRadar, Forbes, G2, Which?, NerdWallet | These domains are trusted by both models across most categories |
External Evidence Supports This Pattern
Search Engine Land and Evertune reviewed about 25,000 of the most-cited URLs and found that across nearly 400 million citations, 63% pointed to listicles, with 71% to 86% of them ranked numbered lists. A separate Search Engine Land analysis of 8,000 citations found citations favour listicles, articles and product pages over homepages.
Profound's citation-pattern research and a 680-million-citation audit found only about 11% of domains are cited by both ChatGPT and Perplexity, close to the 18% overlap here. Guidance from Masset and Wellows' 7,000-query report reach the same conclusion: get onto ranked lists, keep them current, back claims with data.
What This Means for Your Content
Stop optimising the homepage for AI. It is the least-cited page type. Build genuine ranked comparison and best-of pages, structured as numbered lists of named options.
Win the third-party list. Most citations sit on publisher lists you do not own. Treat inclusion in TechRadar, Forbes Advisor, G2 and category equivalents as a PR and outreach target.
Run two playbooks. Match the source mix to where your buyers ask: authority and UK editorial for ChatGPT, breadth and analyst coverage for Claude.
You Cannot Fix What You Cannot See
Because citations concentrate on a knowable set of hubs, the right thing to measure is how often and how prominently you appear inside the sources these models pull from, per model. That is exactly what LLM Scout tracks: citation frequency, citation position and share of voice inside AI answers across ChatGPT, Claude, Gemini and Perplexity.
Try LLM Scout: sign up at llmscout.co/register and use code CITATION for 10% off a yearly subscription (valid until end of July 2026).
Bottom Line
AI does not cite brands, it cites lists. Three in four citations went to ranked comparison pages, and fewer than one in fourteen to a homepage. ChatGPT and Claude agree on format but disagree on sourcing, overlapping on one source or fewer on nine prompts in ten. Win AI visibility by earning placement in ranked comparison content, and measure it per model.
Methodology: LLM Scout, in partnership with citations.press. Dataset: 10,000 prompts across 50 industries, 63,500 modelled citations. Models: ChatGPT, Claude. Page types inferred from URL structure. July 2026.
