AI Citation Studies Compared: Muck Rack, Orbit and Their Limits

The most useful recent data on which sources AI assistants cite is Muck Rack's October 2026 "What Is AI Reading?" report, which says more than 84% of cited links come from earned and third-party sources. Orbit Media, two arXiv papers and one first-party index add more. They agree that ChatGPT, Claude, Gemini and Perplexity cite different sources, and they disagree on Reddit, Wikipedia and Google overlap. This post is current as of 13 November 2026.

Key Takeaways

• Muck Rack's report, updated 28 October 2026, analyzes 40 million AI responses and 207 million citations, and finds about 99% of cited links come from non-paid sources and more than 84% from earned and third-party sources (Muck Rack).
• Agreement between engines is low. Only 212 domains were cited by all four platforms in Muck Rack's September sample of 1,020 prompts, and Orbit Media found the four models named the same domain for the same question in 30 of 1,792 combinations (1.7%) (Orbit Media).
• Repeat runs disagree with themselves. Muck Rack's same-day repeat test found ChatGPT reused 41% of its cited domains, Gemini 45% and Perplexity 94%.
• The studies conflict on Reddit. Orbit saw Reddit at 0% of Claude and ChatGPT citations; the AuthorityTech index saw it in 19.7% of Perplexity runs and none of ChatGPT's. Prompt sets, dates and access methods differ, so this is not a contradiction.
• Being retrievable comes first. An observational arXiv paper found brand mention rates of 2.8% on GPT when neither the brand nor its domain appeared in the retrieval path, against 49.0% with an own-domain citation.
• No study here is a controlled experiment. Treat every figure as a description of one prompt set on one date.

The studies side by side

I read the abstract pages for both arXiv papers, not the full PDFs, which would not extract as text. Anything I say about them is limited to what the abstracts state.

What the studies agree on

Engines cite different things. Muck Rack's 212 shared domains and Orbit's 1.7% agreement point the same way. The AuthorityTech index adds a sharp example: techradar.com appeared in 9.0% of ChatGPT runs but 0.1% of Google AI Mode runs. If you test one assistant, you have not tested the others. My earlier post on why ChatGPT, Gemini and Perplexity differ covers the mechanics.

Rankings and citations are loosely linked. Orbit found 10% to 30% of cited domains also appeared in Google's top 10 for matching queries. At exact-URL level, overlap fell to roughly 1% to 5% for Gemini and ChatGPT, and 13% to 23% for Perplexity. That fits what I wrote in AI Overview citations have decoupled from top-10 rankings, though Orbit's sample is small.

Answers are unstable. Orbit reran the same questions two days apart: Claude kept 30% of its URLs, Gemini 38% of its domains and Perplexity 65%. Muck Rack's same-day test shows the same ordering with Perplexity most stable. A single check of "do we appear" tells you very little.

Mentioning and citing are different events. Orbit measured both. Claude mentioned brands in 70% of answers but cited them in 40%; Gemini 72% against 41%; Perplexity 67% against 44%; ChatGPT 56% against 47%. This is the gap I describe in citations are not recommendations.

Where they conflict, and which I trust more

Reddit. Orbit's latest run had Reddit at 0% of Claude and ChatGPT citations, 3% of Perplexity and 1% of Gemini. AuthorityTech had it at 19.7% of Perplexity runs. The AuthorityTech index itself notes that an earlier Semrush study had Reddit and Wikipedia as ChatGPT's most-cited domains in 2025, and says the differences in surface, timing and question set mean the gap is not necessarily a contradiction. I would not use any of these as a current Reddit share. For the history, see Reddit's ChatGPT citation collapse.

What counts as "earned". Muck Rack lists its categories (journalism, government and NGO, academic, encyclopedic and aggregator, third-party corporate, owned, paid) but does not say how sources were assigned to them. Its own numbers show owned media rising from 13.7% to 15.2% between May and September, and third-party corporate content from 24.0% to 25.1%. The headline of "84% earned" depends on that unpublished classification, so I would quote it as "Muck Rack reports" and not as a fact about the web.

Journalism's share. Muck Rack reports journalism fell to 22% of citations in September from 27% in May, after holding between 25% and 27% since July 2025. For ChatGPT it fell from 28.1% in January to 22%. Muck Rack also says several major publishers dropped to near zero despite allowing OpenAI's search crawler. This is a vendor report with one snapshot month, and I have no second source for the trend.

The crawler-access finding deserves a caveat

Muck Rack reports that among major publishers blocking Claude's search crawler, 81% received almost no Claude citations, against 12% of publishers that allow it. That is a correlation. Publishers who block may also be smaller, paywalled or less relevant to the prompts used. Still, it supports a basic check I recommend in OAI-SearchBot vs GPTBot: make sure the search-side bots you want citations from are not blocked by robots.txt or by your CDN.

The Tannenbaum paper points the same direction from another angle. In 34,960 unbranded prompt observations, with neither the brand nor its own domain in the live retrieval path, the brand was mentioned 2.8% of the time on GPT and 3.8% on Gemini. With an own-domain citation but no branded fan-out query, mention rates rose to 49.0% and 58.4%. With both, they reached 91.4% and 100%. The author calls the model predictive and observational, and the sample comes from projects monitored by one tool, Aiso, so it may not represent all brands.

Methodology limits to check before you reuse any number

1. Prompt set. Orbit used 72 prompts across three brands. A different industry or intent mix would give different domains.
2. Surface. The AuthorityTech index reaches ChatGPT through OpenAI API models, not the consumer app. Orbit used APIs for three engines and an agent-with-search setup for Claude. Neither equals what a person sees in the app.
3. Date. Orbit's data ends 23 August, Muck Rack's prompt test is September, AuthorityTech's release is 29 October. Engines change retrieval behavior often.
4. Classification. Any "earned versus owned" split needs a stated rule. Muck Rack's page does not give one.
5. Repeats. One run per prompt is a coin flip for Claude, Gemini and ChatGPT given the repeat figures above.
6. Author interest. The AuthorityTech index is first-party data from the vendor's founder, the Tannenbaum sample comes from projects monitored by one tool, and Orbit's author tracked his own site as one of the three brands. That does not make them wrong, but it affects what was chosen and published.

What this means for marketers and founders

My read is that the studies support a few cautious actions and rule out several popular ones.

• Build your own prompt set from buyer questions and run it on the surfaces your buyers use. Start with the method in prompt tracking: your first 50-prompt panel. Run each prompt at least three times and report a range.
• Check bot access this week. Look at robots.txt and your CDN rules for the search bots of each assistant. Do this before any content work.
• Invest in third-party mentions. Orbit's gap between mentions and citations, and Muck Rack's earned share, both point to sources outside your site. Third-party consensus explains why that matters for what engines trust.
• Do not chase a domain list. AuthorityTech's own advice is to skip a generic top-domains list and test your buyers' actual questions. I agree, and the data above shows why: a domain that dominates one engine can be near zero in another.
• Leave platform-specific tricks alone until someone shows a repeatable effect with run-to-run variance reported.

What we do not know yet

• Whether Muck Rack's source classification is consistent across platforms and months.
• Why Perplexity is far more stable across repeat runs than ChatGPT or Gemini. Orbit and Muck Rack both observe it; neither explains it.
• How much of the journalism decline is a real shift in retrieval and how much comes from changes in the prompt set between May and September.
• Whether the arXiv findings on AI-generated sources, roughly 16% of cited sources in the Allaham and Diakopoulos audit, hold outside politics, health and environment queries.

FAQ

Which sources do AI assistants cite most?

It depends on the assistant, the prompt set and the date. Muck Rack reports that over 84% of cited links come from earned and third-party sources, and AuthorityTech found techradar.com and youtube.com the top domains in its ChatGPT runs. No single domain list holds across engines.

Do ChatGPT and Perplexity cite the same websites?

Rarely. Muck Rack found only 212 domains cited by all four of ChatGPT, Claude, Gemini and Perplexity in its September sample. Orbit Media found the four models named the same domain for the same question in 1.7% of combinations.

Is Reddit still the most cited source in AI answers?

The recent studies do not agree. Orbit saw 0% Reddit in Claude and ChatGPT citations in its latest run, while AuthorityTech saw Reddit in 19.7% of Perplexity runs. Check your own prompts on the engines your buyers use.

Does ranking in Google's top 10 get you cited by AI assistants?

Only partly. Orbit found 10% to 30% domain-level overlap with Google's top 10, and 1% to 5% at exact-URL level for Gemini and ChatGPT. Ranking still helps retrieval, but it does not guarantee a citation.

Can I trust a single AI citation check?

No. Muck Rack's same-day repeat test shows ChatGPT reusing 41% of cited domains and Gemini 45%. Run each prompt several times and report the spread.

Does blocking AI crawlers reduce citations?

Muck Rack reports that 81% of major publishers blocking Claude's search crawler received almost no Claude citations, against 12% of those allowing it. That is a correlation, not proof of cause, but it is a reason to audit your crawler rules.

Are AI-generated pages being cited by AI search engines?

An arXiv audit of ChatGPT, Copilot, Gemini and Perplexity on 712 politics, health and environment queries found about 16% of cited sources showed evidence of being AI-generated. The sample is limited to those topics.

Want help with this?

I'm Younus Fardeen, an organic growth, SEO and AEO strategist with 4+ years of marketing experience across edtech and startup teams. If you want help building a citation test for your own category that reports ranges instead of single runs, see my work and reach me through the contact form at younusfardeen.com.