ChatGPT Pulls Answers From Four Sources — And Only One Is Fully Yours
ChatGPT's searches equal about one-fifth of Google's volume, Ahrefs finds — and only ~8% of its citations match Bing's or Google's top 10, reshaping how brands should chase visibility.
- ChatGPT web searches equal roughly one-fifth of Google's query volume, per OpenAI's Adam Fry (Nov 2026).
- Only ~8% of ChatGPT citations matched Bing's top 10, and ~8% Google's, in Ahrefs' 15,000-prompt study.
- 'official' appeared in 13.1% of ChatGPT's English searches in September 2026, up from 10.1% in August.
- Partner publishers' citation share more than doubled from 7.5% to 15.8% between March and July 2026.
- ChatGPT cited pages averaged 458 days newer than Google's organic results across 17 million citations.
ChatGPT's searches now add up to roughly one-fifth of Google's query volume, according to OpenAI's consumer product lead Adam Fry — and new research from Ahrefs shows exactly where those answers come from. The study, by Ahrefs' Si Quan Ong, breaks ChatGPT's information supply into four channels: training data, live web search, licensed publisher content, and the user's own context. Each follows different rules, and each rewards different publisher behavior.
The headline finding for SEO teams: ranking on Google or Bing barely predicts whether ChatGPT cites you. In one 2025 study of 15,000 prompts, only about 8% of ChatGPT's in-text citations appeared in Bing's top 10 — and about 8% in Google's top 10 for the same question. A companion study using ChatGPT's own fan-out queries found only 6.82% overlap with Google's top 10, while 83% of cited results didn't appear in Google's results at all. ChatGPT rewrites your prompt into multiple targeted searches (query fan-out), then picks sources itself. Ranking well helps, but it doesn't decide the outcome.
What's in the training data?
OpenAI's last detailed breakdown dates to GPT-3 in 2020: roughly 60% filtered Common Crawl, 22% WebText2, 16% books, and 3% Wikipedia. Since then, the company says only that models were "trained on diverse datasets." Current knowledge cutoffs are April 20, 2026 for GPT-6 Sol (paid plans) and May 18, 2026 for GPT-6 Luna (free). OpenAI excludes paywalled and dark-web sources, filters spam and hate speech, and increasingly uses synthetic data. Site owners can opt out via GPTBot in robots.txt — and GPTBot is the most-blocked AI crawler in Ahrefs' study of ~140 million websites. The trade-off: future models learn less about your brand.
How does search actually work?
In a 12-question test, 9 prompts triggered a live search; only simple definitions came from training alone. A study of 66.6 million free, logged-out responses (August–September 2026) found responses with fan-out queries contained about 1.6 each. Two patterns stood out:
- 6.8% of fan-out queries used the site: operator, and three in four of those targeted brand websites directly.
- The word "official" appeared in 13.1% of English searches in September, up from 10.1% in August — ChatGPT is hunting primary sources, not roundups.
OpenAI names Microsoft as a search partner and runs its own crawler, OAI-SearchBot, whose index Cloudflare says gained "hundreds of millions of new web pages" via a pilot. Whether OpenAI also pulls Google results through SerpApi, as The Information reported in 2025, remains unconfirmed by OpenAI. Separately, cited pages skew fresh: 458 days newer on average than Google's organic results (17 million citations studied). Schema markup made no measurable difference in a test of 1,885 already-cited pages.
Citations also shift constantly. Glen Allsopp tracked 1,300+ prompts daily for four weeks and found only about a quarter of cited pages repeated the next day, though top domains were steadier — changing about 14% day to day, versus 74% for the rest.
Which pages get cited?
Ahrefs pulled 4,000 US prompts from Brand Radar (October 2026) and sorted citations by question type:
- Health: 82% of citations went to government, medical, and academic sites.
- Definitions: 63% to dictionaries and encyclopedias; Cambridge Dictionary alone got 42%.
- Brand questions: 59% to company websites.
- "Best X" queries: review and advice sites led with 27%.
- Shopping: retailers led with 28%; Walmart appeared in 11% of shopping answers.
Web-search feeds dominate: ChatGPT cited 88% of pages retrieved from web search but under 2% from Reddit, YouTube, or academic feeds (1.4 million prompts, February 2026).
Do publisher deals guarantee visibility?
Since 2023, OpenAI has signed deals with more than 20 publishers, including the Associated Press, Axel Springer, Condé Nast, and News Corp. Getty Images joined in June 2026 and hit #9 among ChatGPT's most-cited domains by November. Among 85 partner domains checked, 10 sit in the top 50 most-cited sites and 47 in the top 500. Partner share of citations among the 200 most-cited domains more than doubled from 7.5% in March 2026 to 15.8% in July, then settled at 12.5% in late September. But deals don't guarantee reach: Vogue drew 42,108 citations; fellow partner The Atlantic, just 724. As Ahrefs puts it: "A partnership makes a publisher eligible to show up in ChatGPT, but it doesn't guarantee it."
What should brands do?
Three practical moves emerge from the data. Get talked about where ChatGPT reads — 94% of new Ahrefs mentions in a self-promotion experiment came from answers citing someone else's page. Be the official answer — 83% of brand-question answers cited the brand's own site, 70% first. And allow OAI-SearchBot, which controls search visibility independently of GPTBot.
Watch the "official" search trend and the growing site: share as signals of where OpenAI's retrieval goes next.
via suganthan.com (Original)
More from Olivia Hart
Show full bio
Staff writer covering money, funding and market shifts around Search rankings and algorithm updates, generative engine optimization at SERP Journal.
216 articles