Which Pages AI Tools Cite, Why the First Citation Counts, and Next-Gen AI Models
Nine new studies: product pages win AI citations, brand mentions follow your own website, AI answers can send more than their share of clicks, and next-gen AI models. One task: put buyer facts on one product page.
This week in one line
AI tools now cite product and service pages most, and they mention the brands whose own pages they find.
We added 9 new studies to LLM Watch Research this week. Five are about AI search. Four are about the next kind of AI model. Each note ends with what it means for your leads.
Study of the week: AI tools cite product pages most
LightSite AI counted about 240,300 links that ChatGPT and Perplexity cited from July to November 2026. It compared them with about 207,100 links from October 2025 to March 2026.
- Product and service pages are now the largest group: 41% of ChatGPT citations and 45% of Perplexity citations.
- Help centers, support pages and docs grew fastest: up 121% on ChatGPT and 160% on Perplexity.
- Company blog posts fell 29% on ChatGPT and 46% on Perplexity.
LightSite sells GEO tools. The study counts citations, not clicks or sales.
Why it matters: Your product, service and help pages do more work in AI answers than your blog. Put clear facts, prices and answers on them.
AI tools mention brands whose own pages they find
Benjamin Tannenbaum studied 34,960 GPT and Gemini runs across 75 projects from June to September 2026. The brand's own website made the largest difference.
- Neither signal
- 2.8%
- Own site cited
- 49.0%
- Brand search only
- 64.4%
- Both signals
- 91.4%
A brand that the AI mentioned in one run was also more likely to get a mention in the next run. The author works for a GEO company.
Why it matters: If AI tools do not find your own pages, they rarely mention you. Make sure your website answers the questions that buyers ask.
AI answers sent one site 21% of its Google clicks
Cyrus Shepard of Zyppy removed his site from Google AI features for 28 days in September, then added it back. AI Overviews and AI Mode gave the site about 10% of its impressions but about 21% of its clicks.
The AI click-through rate was about 0.9%, against about 0.4% for the site in all of Google Search. This is one site, so use it as a direction, not a rule.
Why it matters: A citation in Google's AI answers can send more than its share of clicks. Stay in Google's AI features and work to get cited.
Subscribe to A/GEO Research
Get one short email each week: new AI models, new studies on AI search, and one task for your website. Each note says what it means for your leads.
In a chat, the first citation counts most
Two new papers look past a single AI answer.
- Chats: a source that an AI cites early in a chat is more likely to get cited again later. In the authors' model, this effect passes 2x within ten turns. The numbers come from a model, not from live AI tools. (Yu, Zhou, Yang, Ding and Sato, HAI '26, September 30, 2026)
- Buyer intent: the same page rewrite helps some buyers more than others. Across 4 domains and 10 buyer intents, the gain changed with the buyer's need and stage. (Lei, Zhang, Xiao and Wei, November 3, 2026)
Why it matters: Write a clear page for each buyer stage. A citation in the first answer can stay for the full chat.
Next-gen AI models: faster drafts and longer memory
Most AI models today write one word piece at a time, from left to right. Diffusion models write many word pieces at once, then improve the draft.
- About 6x faster research: a deep research agent built on a diffusion model ran about 6 times faster on 100 tasks, with about the same report quality. (Gazzaev, Gavrilov, Mozikov and Muravyov, EMNLP 2026 Industry Track, October 20, 2026)
- Self-correction: a small checker tells a diffusion model when to rewrite a part of its draft. Accuracy went from 35.3% to 43.4% on five reasoning tests. (Oh, Kim, Koo, Kim, Seo, Seol and Lee, November 12, 2026)
- Cheaper training: looped diffusion models reached the same loss with up to 3.3 times less training compute. (Utpala, Lai, Mitsufuji and Vepakomma, October 21, 2026)
- Longer memory: one shared memory for all layers scored up to 2.6 points higher on a long-context test. These were small test models. (Zefan Cai et al., October 4, 2026)
Why it matters: Future AI agents will read and compare more pages for each question, and faster. Pages with clear facts that are easy to quote will be easier for them to use.
New models this week
- Anthropic: Claude Fable 5.2 and Claude Mythos 5.2 (November 9).
- Meta: Muse Ember (November 9).
- OpenAI: a beta for businesses in ChatGPT, with verified business profiles (November 10).
- Google: Gemini 4 Carbon opened to paid API and AI Ultra users (November 10), Gemini Decision 1 (November 11), and Gemini 4 Hydrogen (November 12). Gemini 4 Hydrogen is now the default model in AI Mode.
- Qwen: Qwen4-Max-Preview (November 11).
Why it matters: Google AI Mode now runs on a new default model. Ask your top buyer questions again this week.
This week's task: put buyer facts on one product page
- Pick the product or service page that brings in the most sales.
- Write the 3 questions that buyers ask before they buy. Use each question as a heading on that page.
- Answer each question in 40 to 60 words. Include a price or price range, who it is for and where you work.
- Ask an AI tool your top buyer question. Write down which page it cites. Ask again next week and compare.
First, make sure AI tools can read your site. The free AI Visibility Score runs 37 checks in about 20 seconds and shows what to fix first.
Field notes: LLM Watch has a Next-gen models tab
LLM Watch Research now lists 26 studies. A new Next-gen models tab shows research on new kinds of AI models: diffusion models, self-correction and memory. Each card says what the finding means for your content.
See all 26 studies in LLM Watch
Sources
- Stas Levitan, AI Citation Sources in 2026: What ChatGPT and Perplexity Cite Most, LightSite AI, November 3, 2026
- Benjamin Tannenbaum, From Prompt to Recommendation: A Fitted Stage Model of Brand Visibility in AI Search, arXiv 2609.23162, September 19, 2026
- Matt G. Southern, Google's AI Features Sent One Site An Estimated 21% Of Its Clicks, Search Engine Journal, November 9, 2026 (Cyrus Shepard, Zyppy)
- Yu, Zhou, Yang, Ding and Sato, Conversational Capture: A Trajectory-Level Framework for Evaluating Generative Engine Optimization in Multi-turn Human-Agent Interaction, arXiv 2609.40069, September 30, 2026
- Lei, Zhang, Xiao and Wei, IntentGEO: Intent-aware Generative Engine Optimization for Recommendation, arXiv 2611.04225, November 3, 2026
- Gazzaev, Gavrilov, Mozikov and Muravyov, DiffResearch: Accelerating Deep Research with Diffusion Language Models, arXiv 2610.28909, October 20, 2026
- Oh, Kim, Koo, Kim, Seo, Seol and Lee, Learning When Diffusion Language Models Should Regenerate, arXiv 2611.17320, November 12, 2026
- Utpala, Lai, Mitsufuji and Vepakomma, Looped Uniform Diffusion Language Models, arXiv 2610.30066, October 21, 2026
- Zefan Cai et al., Universal Test-Time Training, arXiv 2610.05484, October 4, 2026
- A/GEO Research, LLM Watch research list and release log, checked November 15, 2026