Which sources AI models cite on enterprise AI questions: 240 answers checked

The 240 answers cited 893 sites between them, counting each site once an answer. The vendors' own sites came first, with openai.com in 92 answers, claude.com in 46 and microsoft.com in 33.

Fourth was reddit.com, in 31 answers, and all 31 were from Gemini 3.5 Flash-Lite. Past the top 11 the list thins out fast: 311 of the 417 different sites were cited in only one answer. The top 10 sites made up 286 of those 893, or 32.0%.

Each answer counts once per site, whichever of the site's pages it cited. Subdomains are grouped, so help.openai.com counts as openai.com.

189 of the 240 answers cited at least one product, help or documentation site of an AI vendor: OpenAI, Anthropic, Microsoft, Google or Amazon. GPT-6 Luna did so in 109 of its 120 answers and Gemini 3.5 Flash-Lite in 80 of 120.

OpenAI's sites appeared in 95 answers, Anthropic's in 70 and Microsoft's in 35. Google's own sites appeared in 12 and Amazon's in 5. 83 of the 120 questions named an OpenAI or Anthropic product, against 15 naming a Microsoft product and 5 a Google one.

A vendor's page is the place to check that vendor's prices, limits and data terms. Read the date on the page and the plan it applies to, because the vendor is describing its own product.

GPT-6 Luna cited few sites and mostly vendors. It averaged 1.37 sites an answer, drew on 33 different sites across its 120 answers, and cited no forum or video site at all.

Gemini 3.5 Flash-Lite cited more, and more widely. It averaged 6.08 sites an answer and drew on 399 different sites. In 50 of its 120 answers it cited reddit.com, youtube.com, medium.com or another user-generated site. Eleven of its answers carried no citation. The run let Gemini decide whether to search, while OpenAI's requests had to search.

The two models cited at least one site in common on 63 of the 120 questions, and two or more on only 4. The shared site was openai.com on 35 of those 63.

Fifteen questions asked about data residency, audit logs or whether a vendor trains on company data. 27 of those 30 answers cited a vendor's own site. Two cited a government site, rulebook.sama.gov.sa and ncsa.gov.qa, both on the Gulf questions.

Across all 240 answers, 14 cited a government domain. 6 of those were on the questions about time saved and return, among them the UK government's Copilot trial. No answer cited the ICO or any europa.eu site. This is not legal advice. For the rules that apply to you, read the regulator's own guidance.

The topic groups are 1AYM's, made from the topics the questions were written for. Every question on audit logs or training on company data sits in the first group. Each question's group is in the CSV.

Open the cited page and check that its address is the vendor's own domain. 20 of the 240 answers cited a site with ChatGPT, Claude, Codex or Cowork in its address. None of those sites is on the vendor list in the method below.

Check the date and the plan on that page. A price or limit from last quarter, or for another plan, answers a different question. For what each vendor's own pages say about the two main work plans, see our ChatGPT Enterprise and Claude Enterprise comparison.

Ask the same question of a second model. On 57 of the 120 questions the two models cited no site in common. One leaned on vendor pages and the other mixed them with forums and videos, so a second answer shows quickly where the sources disagree.

On 29 September 2026 a script sent each of 120 questions once to OpenAI's Responses API, model gpt-6-luna. Web search was required and the location set to the US. It sent each question once to Google's Gemini API, model gemini-3.5-flash-lite, with Grounding with Google Search. All 240 answers completed and none failed. These are API answers with those settings, and the ChatGPT and Gemini apps may cite differently.

The questions are a purposive sample. 1AYM wrote them from its own list of enterprise AI buyer topics, three for each of 40 topics, and 83 of them name an OpenAI or Anthropic product.

A citation is a link the answer itself carries. Each answer counts a site once, and subdomains are grouped to the site. Vendor sites are the product, help and documentation sites of OpenAI, Anthropic, Microsoft, Google and Amazon, listed in full in the notes for engineers. User-generated sites are reddit.com, youtube.com, medium.com, substack.com, dev.to and facebook.com.

The full data is one CSV with a row for each answer: the question, its topic, the model and the domains it cited, as recorded. It is linked under the figure below and in the sources. Cite this page and the run date of 29 September 2026.

If your team is choosing an AI assistant and much of its research so far is AI answers, bring in 1AYM. Our AI Opportunity & Feasibility Sprint runs a feasibility check against your data and systems and settles build versus buy. It ends in a roadmap with a governance review, at a fixed scope over typically two to four weeks.

Rolling out AI tools to a workforce is core 1AYM work. We roll out the tool you choose with your operations and technical teams, and train your staff at every technical level. We have rolled out ChatGPT, and we have rolled out Claude Enterprise for a client. We are model-selective across OpenAI, Anthropic, Gemini and open-source, and fixed-scope work can start within a day of the scope being signed. Bring the shortlist and the pages behind it to a 30-minute call, or email enquiries@1aym.com.