Four of the five AI models we measured mostly cite the company blog. ChatGPT does not: its most cited category is product pages, at 31.3%, and its blog drops to 12.6%. On product, though, it nearly matches Claude (28.2%) and Perplexity (30.6%). Add guides, PDFs, documentation and pricing and 62% of what ChatGPT cites is product material. Shared ground exists. The hard cut is blog, PDFs and social.
All 20 pages, with every table and all 26 references.
Blog vs product, model by model
Share of citations to the company blog and to product pages, out of each model's total. Violet is blog. Gray is product. In bold, the category each model cites most.
You do not have to pick Claude or ChatGPT for everything. Start with product pages that answer the question. Then specialize: more blog for Claude, more product, guides and PDFs for ChatGPT, social and video for Google AI Mode.
What each model cites, one by one
Each card orders categories by volume. Under each one: average position (longer bar = better place in the list) and the share of times that citation reaches first place. They do not add up to 100% because editorial notes, comparisons, documentation, pricing, case studies and a 2% unclassified share are left out. ChatGPT and Claude get the detail. The other three, shorter.
ChatGPT
2,151 citationsThe outlier: the only model whose top category is not the blog. Product leads on volume (31.3%). Guide, blog and product average almost the same place (between 2.95 and 3.78), so do not trade volume for half a slot. What does separate is first place: guides hit the top 28.1% of the time, versus 16.9% for product. It also cites the most PDFs and the most government sources (9.6% vs 0.7% for Gemini). Write product pages and whitepapers to answer, not only to convert.
Claude
2,909 citationsThe most editorial: 92.3% of citations go to company sites. ZipTie reports blogs in 43.8% of its primary citations, almost the same as ours (42.5%). The blog wins volume and loses place: only 10.6% reach first, versus 23.3% for guides and 22% for the homepage. Here the article buys mentions; guides and home fight for the front seat.
Gemini
1,567 citationsThe heaviest homepage citer (16.5%) and the least deep (58.5% to an internal page). The home is also best placed (position 7.37; 12% first). The blog leads volume and sits back in the list (15.10; 5.6% first). If your home answers the category question, it matters here.
Perplexity
1,477 citationsThe most balanced on volume: blog 33.1% and product 30.6%, nearly tied on average position too. Guides reach first place 14% of the time, a bit more than blog or product. A good place to start if you can only measure one model.
Google AI Mode
2,514 citationsThe only one that cites social and video in earnest: 12.1% of volume, and better placed than the blog (position 7.29 and 9.8% first, versus 12.35 and 4.6% for the blog). Claude adds up to 0.5% across social and video. A twenty-fourfold gap. For AI Mode a video or a post can be the source. For the others, almost never.
What kind of site each citation comes from
Claude sends 92.3% of its citations to company sites. ChatGPT puts 9.6% on government, far ahead of the rest. Google AI Mode is the only one with social and video in double digits: 12.1%. That 76% to 92% from “company” sounds good until you read it properly: it is the sum of every company in the world, your competitors included. Across the categories we analyzed, no brand domain exceeded 10% of the citations in its own category.
Read the full analysis in the study, section 04
How similar two models really are
To compare two lists of sources we use the Jaccard index. The idea is simple and needs no formula: take every distinct site the two models cited between them, and look at what share both of them cited. It returns 0 if they share nothing and 1 if they cite exactly the same thing.
Any combination of two AI models we compare lands around 0.104 at the site level. In plain terms: out of every 100 sites cited between the two, about 10 are cited by both and the other 90 by only one. And when we compare the exact page instead of the site, the number drops to 0.035, that is 3 or 4 out of every 100. They may trust the same sites, but they rarely pick the same page inside those sites.
This is not a quirk of our sample
The thesis that every model cites differently is the most replicated finding in the field and the one budgets ignore most. Three teams with samples far larger than ours land in the same place by different routes.
| Who | Sample | What they found |
|---|---|---|
| Writesonic | 161,286 queries | 17% of sources shared, 3.8% universal |
| Profound | 100,000 queries | 89% distinct citations between ChatGPT and Perplexity |
| Averi | 680 million citations | 11% domain overlap between those two |
The work of Chen, Wang, Chen and Koudas, at the University of Toronto, documents the same thing from another angle: these services differ from each other in domain diversity, in freshness of information, in stability across languages and in sensitivity to how a question is phrased. What our analysis adds is not the thesis, it is the resolution: they do not just cite different things, they cite different mixes of page types, and that can be acted on.
With the methodology, the tables by industry, and what the study does not say.
What to do with this
Three decisions, and all three come before any content calendar.
Start with what is shared, then specialize in what is left. Product pages already weigh between 22% and 31% across the five: Claude, ChatGPT and Perplexity almost touch there. The same exact URL rarely wins everywhere (3.5% overlap). Where it pays to pick a model is the rest: more blog for Claude, more product, guides and PDFs for ChatGPT, social for AI Mode.
Stop treating the product page as sales material. It is between 22% and 31% of what any of the five cites, and it is the category ChatGPT cites most. Today most of those pages are written to convert someone who already decided. To be cited, they have to answer the question first.
Split your Google dashboard from the rest. Google AI Mode cites 12.1% social and video and Claude 0.5%. If you look at both through one averaged figure, that average describes neither.
All of this comes from the same work that presents the study's eight findings. The way to know which model you show up in is to measure it model by model: AI Visibility (weekly scan of more than 100 questions) or GEO Observer (your real Chrome sessions). Both can be tried with no card.