What AI assistants cite: 32,000 sources from 7,400 answers
When a buyer asks ChatGPT, Gemini or Google AI Mode which product or provider to choose, the answer is built from web pages. We collected the sources of 7,911 such answers, to 2,650 buyer questions about 221 brands in B2B software, retail, services and consumer goods, and sorted the 34,588 citations by site and by kind of page. The usual GEO advice (get on Reddit, collect reviews on G2) does not match what the answers cite.
Key findings
- Brand websites are the first source. 33% of citations go to the sites of brands: 28% to competitors and other brands, 5.5% to the brand the question is about. On ChatGPT, 37%.
- There is no must-have site. The answers cite 13,982 different sites, and 64% of them appear in a single answer. The 10 sites cited most account for 3.3% of citations, the top 100 for 13%.
- Reddit and review platforms weigh little. Reddit is cited in 1.4% of answers (3.0% in English-speaking markets), review and rating platforms in 6.6%. On B2B software questions, G2, Capterra and their peers appear in 74 of 2,602 answers (2.8%).
- Comparison pages are a fifth of what gets cited. On brand, media and other websites, 23% of cited pages are rankings, comparisons or reviews (29% on Gemini). Pricing pages: 1.4%.
- The three models read different webs. On the same question, ChatGPT and Gemini share 8.3% of their cited sites on average, and none at all in 52% of cases.
- Answers that cite a brand's site name the brand. 97% of them do, against 18% of the answers that don't cite it.
Brand websites first, a long tail of niche sites second
We sorted every cited site into a kind: the brand the question is about, its competitors and other brands, review platforms, Reddit and forums, social and video, marketplaces, news, public and reference sites, and everything else. A site counts once per answer, however many of its pages are cited.
A site counts once per answer. Competitors and other brands: the sites of the competitors tracked in each audit and of the other audited brands.
Two kinds dominate. Brand websites take 33% of citations, and this is a floor: we only recognise the brands tracked in our audits, so the sites of other vendors land in the last group. That last group, other websites, takes 56%: specialist comparison sites, trade blogs, consultants, local businesses, niche publishers. Everything the GEO playbooks usually put first (review platforms, Reddit, YouTube, news, Wikipedia) adds up to less than one citation in ten.
For a brand, this means two things. Its own site and its competitors' sites are where the models read how a category is described, so a vendor whose pages don't state who it serves and how it compares leaves that description to its rivals. And the rest is spread across thousands of small sites, specific to each sector and each question: the sources that matter have to be found question by question, not taken from a generic list.
No site is in more than a few percent of answers
The site cited most, YouTube, appears in 3.5% of answers. Below it come Google Maps (on local questions), vendors with large content libraries (HubSpot, Pennylane, Cegid, Pipedrive, Salesforce, Sage), Reddit, LinkedIn, a French comparison site for software (lafabriquedunet.fr), public sites and two review platforms.
None of these is a gate: the top 15 together are in a small minority of answers. What decides whether a brand is named is what the pages retrieved for its own questions say, and those differ from one sector and one question to the next.
Reddit, G2 and Wikipedia: smaller than their reputation
Review and rating platforms: Trustpilot, Gartner, Which?, Tripadvisor, G2, Capterra and their peers. G2 and peers: G2, Capterra, GetApp, TrustRadius, Software Advice and Appvizer, on the questions about B2B software.
Reddit is cited in 1.4% of all answers. It does better in English-speaking markets (3.0%) than in France (0.9%), and almost only on Google AI Mode and Gemini: ChatGPT barely cites it in our data. Wikipedia is cited in 31 answers out of 7,911. Review and rating platforms (Trustpilot, Gartner, Which?, Tripadvisor, G2, Capterra…) appear in 6.6% of answers.
The gap is widest for software. On the 2,602 answers about B2B software, G2, Capterra, GetApp, TrustRadius, Software Advice and Appvizer together appear in 2.8%. Reviews on these platforms may still shape what models learned in training, and they matter to buyers who visit them. But as cited sources of live answers, they are a minor input: a smaller vendor shut out of their rankings is not shut out of AI answers.
Comparison pages, home pages and articles
Read from the address and title of each page: "best", "top 10", "vs", "alternatives", "review", "comparatif"… for comparisons; "how to", "guide", "what is"… for guides.
On the sites that publish their own content (brands, media, other websites), 23% of the pages cited are comparisons, rankings or reviews: "best tools for…", "top 10…", "X vs Y", "alternatives to…". The share rises to 29% on Gemini. Home pages make up 24%, mostly on ChatGPT (34% of its citations), which links the brands it names to their sites. Pricing pages are rare (1.4%), even though buyers ask about prices: most pricing pages don't state a price a model can quote.
Comparison pages are written by vendors as much as by publishers. When a vendor publishes "the best tools for X", with itself and its competitors described fairly, that page becomes a source for the question it answers.
ChatGPT, Gemini and Google AI Mode don't read the same web
| Model | Answers | Sites cited per answer | Answers without a source | Citations to brand sites | Answers citing the brand's own site | Comparison pages |
|---|---|---|---|---|---|---|
| ChatGPT | 2,635 | 6.4 | 0.3% | 37% | 31% | 20% |
| Google AI Mode | 2,627 | 4.2 | 3.6% | 33% | 25% | 26% |
| Gemini | 2,649 | 2.5 | 33% | 24% | 12% | 29% |
ChatGPT cites the most sites (6.4 per answer), Google AI Mode 4.2, Gemini 2.5, and Gemini gives no source at all in 33% of its answers. ChatGPT also cites the brand's own site most often: in 31% of its answers, against 25% for Google AI Mode and 12% for Gemini.
The bigger difference is which sites. For each question, we compared the sites cited by two models:
| Models compared | Questions where both cite sources | Sites in common, on average | Questions with no site in common |
|---|---|---|---|
| ChatGPT and Google AI Mode | 2,514 | 13% | 34% |
| ChatGPT and Gemini | 1,751 | 8.3% | 52% |
| Google AI Mode and Gemini | 1,673 | 13% | 40% |
On average, ChatGPT and Gemini share 8.3% of their cited sites on a given question, and have none in common in 52% of questions. Google AI Mode overlaps a little more with each. Optimising for "AI" as one channel misses this: a page that ChatGPT retrieves for a question may never reach Gemini's answer to the same question.
Being the source and being named go together
When an answer cites the brand's own site, it names the brand in 97% of cases. When it doesn't, the brand is named in 18%. This is a correlation, not proof that a citation causes a mention: a model that has decided to name a brand often links to its site. But it shows how closely the two are tied, and it is the brand's own pages that the models quote when they describe it.
Local questions, global sources
On the questions asked in France, in French, only 36% of the citations point to a .fr site. The rest goes to .com sites, many of them international vendors and publishers. A brand that only publishes in one language, or only on its local domain, competes with English-language sources even on local questions.
What this means for brands
- Start with your own site. It is the source the models quote when they describe you. Say who you serve, what you do better and what you cost, in plain sentences a model can lift.
- Publish the comparisons your buyers search for. Rankings, "vs" pages and alternatives pages are a fifth of what gets cited. If you don't write a fair comparison of your category, a competitor or a publisher will.
- Find your sources question by question. With 13,982 sites cited and most of them once, generic lists of "sites to be on" don't work. Look at the pages cited for your questions, in your market, and get listed or quoted there.
- Don't bet everything on review platforms and Reddit. They are worth having, for buyers and for training data, but they are a small share of live citations.
- Track each model separately. ChatGPT, Gemini and Google AI Mode cite different sites for the same question: a gain on one says little about the others.
Method
The data comes from 221 audits run by Rempa between October 12, 2026 and October 28, 2026: 162 brands in France, 42 in the UK, the US, Ireland and Australia, and 17 in other markets; 73 of them B2B software vendors. For each brand, 12 questions written as its buyers would ask them (without naming the brand) were asked once to ChatGPT, Gemini and Google AI Mode, from the brand's country and in its language, in the consumer versions of the assistants: 7,911 answers. See how Rempa collects and reads AI answers.
The answers gave 48,914 source links. We removed the links a model gives to itself (Google search pages, chatgpt.com, image caches) and counted each site once per answer: 34,588 citations. A site is its registrable domain (shop.example.co.uk is example.co.uk), except Google Maps and the app stores. The brand's own site and its competitors' come from each audit; review platforms, forums, social networks, marketplaces, public and reference sites and news sites from lists and patterns we wrote, applied to the domain. The kind of page is read from its address and title.
Limits: the brands were chosen as prospects, mostly mid-sized and mostly in France, so the shares reflect that mix. Each question was asked once per model, and AI answers vary from one session to the next. Sorting by domain is approximate: a news site we did not list counts as "other websites", and a brand we did not track as well. The shares are robust at this scale; small differences between two kinds of site are not.