AI answers lean on third-party sources far more than on a company’s own pages: Muck Rack found that 84% of more than 25 million citations from ChatGPT, Claude and Gemini were earned media, with owned and paid content sharing the other 16% (Muck Rack, May 2026). The figure moves with the engine and the question, and a Yext study of 6.8 million citations put first-party websites at 44% for the right kind of query (Yext, October 2025).
Both numbers can be true at once, and a CEO deciding where to spend the next content dollar needs to know why. The short version is that your blog is raw material for the answer, and other people’s pages are what the answer is built from.
Do AI answers trust a company’s own blog or third-party sources more?
Third-party sources, by a wide margin, on the questions buyers ask when they are comparing options. Muck Rack analysed more than 25 million links across 17 industries and sorted them by who produced the page. Journalism alone made up 27% of citations. Press releases, the format most companies use to speak for themselves, made up 1.1% (Muck Rack, May 2026).
A University of Toronto team measured the same preference from the other side. Mahe Chen, Xiaoxuan Wang, Kaiwen Chen and Nick Koudas compared Google with AI engines across consumer electronics, automotive and banking queries and described a “systematic and overwhelming bias” toward earned media over brand-owned content (University of Toronto, September 2025). For well-known brands, ChatGPT drew 93.5% of its citations from earned sources and 6.5% from brand-owned ones. For niche brands the split was 95.1% to 4.9%.
The engines differ. Gemini took 25.1% of its citations for well-known brands from brand-owned pages, nearly four times ChatGPT’s share, and Perplexity sat at 8.8% (University of Toronto, September 2025). A company that checks only one assistant will draw a conclusion that holds for that assistant alone.
Chen and colleagues did not run the experiment to embarrass anyone’s content team. They were describing how a system behaves when it must justify an answer to a person who cannot check it. A model that repeats a company’s description of itself carries the company’s risk. A model that repeats what a trade publication, a review site and an analyst all say shares that risk across several parties, and the engineers behind these products have good reason to prefer the second pattern.
Why do the studies disagree about how much first-party content gets cited?
Yext reached the opposite headline. Its analysis of 6.8 million citations from more than 1.6 million AI responses across Gemini, OpenAI and Perplexity found that 86% came from sources a brand can manage or strongly influence: 44% from websites, 42% from business listings, 8% from reviews and social, and 6% from sources outside a brand’s reach (Yext, October 2025).
The steelman for that result is a fair one. Many questions put to an AI are local and factual, such as opening hours, locations and whether a product is stocked. For those, the company’s own page is the authority, and a model would be foolish to cite a journalist instead. Yext also sells listings and reputation software, which is worth knowing when reading its conclusions, though the counting method is disclosed and the sample is large.
The studies also define their terms differently. Yext counted business listings as brand-managed, a bucket the Toronto and Muck Rack work would likely file under third-party. Toronto and Muck Rack tested questions about which product or provider to choose, where the buyer wants a judgement rather than a fact. My reading is that the gap comes from definitions and question types, and no study contradicts another once those are held still. A comparison question favours the outside voice and a lookup question favours the owner.
A mid-market B2B company lives mostly on the comparison side of that line. Nobody asks an assistant for your opening hours before a seven-figure software purchase. They ask which vendors are credible, what customers say and how the options differ, and those are the questions where earned sources dominate. The earlier post on G2 and Capterra listings covers one version of that pattern.
Your blog tells the machine what you claim. Other people tell it whether to believe you.
Being cited also pays less than it used to. Pew Research Center tracked 900 U.S. adults and 68,879 Google searches in March 2025 and found that users clicked a result link in 8% of visits when an AI summary appeared and 15% when it did not. Links inside the summary itself drew clicks in 1% of visits (Pew Research Center, July 2025). The answer often finishes the job, so the reputation it repeats matters more than the traffic it sends.
What can a CEO do about it this quarter?
Do not commission a larger blog calendar. The data says volume on your own domain is the weakest lever for comparison questions. Begin by finding out what the assistants currently say and which outside pages they lean on. Each move below needs a laptop, an hour or a conversation, and no marketing training.
Ask the machine ten buyer questions
Type the ten questions a prospect would ask about your category into ChatGPT, Gemini and Claude. Save the answers and the linked sources in one document. It takes an afternoon and costs nothing.
List the domains that get cited instead of you
Tally which outside sites appear across those answers. The five that repeat are your real distribution list, whatever your editorial calendar says.
Check the listings you already own
Yext counted 42% of citations from listings. Have someone confirm that your profiles on review sites, directories and partner pages state the same name, category and pricing as your website.
Hand one outside voice a quotable fact
Choose one analyst, trade editor or customer and give them a specific number from your own data that they can print. A named figure travels further than a press release, which made up 1.1% of citations.
The part nobody wants to hear
None of this is hard to grasp. A leadership team can read the Muck Rack figure on Monday and agree by lunch that outside credibility matters. The gap sits between knowing and doing, and it is structural. Earning a mention from someone else means a long, uncertain conversation with a person who owes you nothing, and a blog post is something a team can finish this week and report as done. Work that can be finished beats work that has to be earned in almost every internal ranking.
Most companies also hold plenty of tools and no operating system. There is a monitoring dashboard, a content platform and a few agencies, and nothing that connects what the assistants say about the company to who will do something about it. Strategy belongs to whoever has ten free hours a week, which is no one, so the output defaults to the blog because the blog is the thing the company controls.
A company that senses what machines and buyers are saying, adapts what it does on that evidence and keeps a human accountable for the result will hear about the problem within a quarter. Most will keep publishing to the one source the data says carries the least weight, and the next study will say so again. If someone needs to own the link between those outside voices and your own pages, that is the work a fractional CMO is built for.
AI answers repeat what outside sources say about you, so your own blog earns its keep by giving them something worth repeating.
This post is part of AI search visibility: every question, in the order it comes up, a reading guide to the whole series.