Most AI citations come from pages you don't own. Your own pages still have a job.

Cloud AI cites who it can see. A new study adds the uncomfortable part: most of what it sees about you isn't on your website.

On October 28, 2026, Muck Rack and its Generative Pulse team released the October 2026 edition of What Is AI Reading?, the fourth round of their study on how AI assistants cite sources. It draws on Generative Pulse data covering more than 40 million AI responses and 207 million citations in 2026, plus a fresh September run of 1,020 unbranded consumer prompts across ChatGPT, Claude, Gemini and Perplexity.

The short version

• Most citations aren't yours. More than 84% of AI citations came from earned and third-party sources: journalism, Wikipedia and other encyclopedic sites, government and nonprofit sites, academic research, and other companies' pages. That share has ranged from 82% to 89% across all four editions.
• Paying doesn't buy much. Paid and advertorial content was 0.3% of citations, the same as in May. About 99% of cited links came from non-paid sources.
• Your own pages still count. Owned media rose from 13.7% to 15.2% of citations. Muck Rack says product pages and practical guides are frequently cited for product and how-to questions.
• The engines don't agree. Only 212 domains were cited by all four AI providers.
• One answer isn't a measurement. Run the same prompt twice in one day and ChatGPT repeated just 41% of its cited domains, Gemini 45%, Perplexity 94%.
• Crawler access is a precondition. 81% of major publishers blocking Claude's search crawler got almost no Claude citations, compared with 12% of those allowing it.

What this means if you run a business

There are two kinds of pages AI can cite about you: the ones you own, and the ones other people write. The study says the second kind does most of the work. That doesn't make your own site optional. It changes what your site is for.

Your own pages win the questions only you can answer. Muck Rack's line is that product pages and practical guides get cited for product and how-to questions. For a local business that means plain pages for each service, your service area, hours, prices if you publish them, and the how-to answers you already give customers on the phone. If those pages don't exist, or say something different from your Google Business Profile, there is nothing first-party for AI to pick.

Other people's pages win the "who's best" questions. When someone asks an assistant to compare options, it leans on sources that aren't yours: directories, reviews, local news, associations, government lists. You can't write those. You can earn them, and you can make sure what they say matches what you say.

Ads don't fill the gap. At 0.3% of citations, paid and advertorial content barely shows up. In ChatGPT's Forbes citations, Muck Rack says editorial content outperformed the pay-to-play Councils program by about 1.7 to 1. Our earlier Field Guide on ChatGPT ads makes the same point from the other side.

Conditions you control

1. Let the search crawlers in. The 81% vs 12% Claude finding is about major publishers, not small businesses, but the condition is the same: if the search crawler can't read you, you can't be cited. Anthropic's crawler page says disabling Claude-SearchBot "may reduce your site's visibility and accuracy in user search results." OpenAI's crawler documentation says sites opted out of OAI-SearchBot "will not be shown in ChatGPT search answers, though can still appear as navigational links." Check your robots.txt and your firewall or CDN rules. Our robots.txt Field Guide shows which token does what.
2. Publish the first-party answer. One page per service. A how-to page for the questions customers actually ask. Same business name, same services, same town everywhere.
3. Earn the third-party mentions. Local news, chambers and associations, the directories and review sites that show up when you search your own category. Mentions on many different sites, not one blurb pasted fifty times.
4. Make the third-party pages agree with you. Wrong hours or an old service list on a directory is a page AI can cite. Fix the ones that disagree.
5. Check more than one assistant. With only 212 domains shared by all four, a ChatGPT check tells you about ChatGPT.

A citation-source check (about 20 minutes)

1. Write three questions a customer would ask: one "how do I…", one "who's the best… in [town]", one about your service or product. Don't use your name.
2. Ask each one in ChatGPT, Gemini, Claude and Perplexity. Note every site cited.
3. Sort the citations into two piles: pages you own, pages you don't.
4. For the "how do I" question, ask: is there a page on your site that answers it? If not, that's the first page to write.
5. For the "who's best" question, list the third-party sites that keep showing up. Those are where your profile, reviews and mentions need to be accurate.
6. Run the same three questions again tomorrow. If the citations change, that's the study's finding, not a fluke. Monitoring tells you what changed. It doesn't fix the page.

What this is not (honesty block)

• Not from OpenAI, Anthropic, Google or Perplexity. None of them has published how they choose which sources to cite. These are observed citations, not documented rules.
• A vendor study. Muck Rack sells PR software and Generative Pulse sells GEO tracking, and its CEO says communicators are "best positioned to lead on GEO." The full report and methodology are behind a download form at generativepulse.ai/report; we worked from the release, not the full method.
• Not a local-business study. The prompts were unbranded consumer questions (the report page's examples are meal kits and travel rewards cards). The Claude crawler figure is about major publishers. Rates for "best roofer in Provo" are unknown.
• Not cause and effect. The study shows where citations land. It doesn't prove that changing your mix of pages changes your citations, or by how much.
• Not stable. Muck Rack says citation behavior "may shift as models are updated or retrained." The 41% / 45% / 94% repeat rates are from one same-day re-run.
• Categories are the study's. What counts as "earned," "owned" or "third-party corporate" is Muck Rack's classification.

Related New Reward Field Guides

• ChatGPT searches for the names it already knows. Make sure it knows yours.
• GEO measurement is noisy, so build a proof loop
• ChatGPT's share of AI prompts fell. Visibility work still has to cover Gemini and Claude.
• robots.txt is a door, not a usage policy

Want the fast read first?

The free AI-visibility score is a quick check: email only, no credit card, a 0–100 read across the engines we track. It won't earn you a mention. It will show whether AI can see your business clearly.