# ChatGPT runs on four feeds, not one brain _Ahrefs shows ChatGPT mixes training data, live search, licensed publishers and your context, with search driving most answers._ **Published:** 2026-11-12 **Source:** https://www.startuphub.ai/ai-news/artificial-intelligence/2026/chatgpt-runs-on-four-feeds-not-one-brain --- [Ahrefs](https://ahrefs.com/blog/where-does-chatgpt-get-its-information/) pulled ChatGPT apart into four information feeds and showed when each one fires. Training is the base layer. [OpenAI](https://www.startuphub.ai/startups/openai) says it uses public web data, partner data and human or synthetic output, per Ahrefs. The last detailed mix was GPT-3 in 2020, about 60% filtered Common Crawl, 22% WebText2, 16% books and 3% Wikipedia. Current cutoffs are for GPT-6 Sol and for GPT-6 Luna, so anything newer needs search. That cutoff is already moving. OpenAI says [GPT-6.1 Sol](https://www.startuphub.ai/ai-news/artificial-intelligence/2026/devrev-pushes-self-service-with-gpt-6-1-sol) offers near-Astra intelligence at one-fifth of Astra prices, per its [launch post](https://openai.com/index/introducing-gpt-6-1-sol/). Search does the heavy lifting for fresh facts. Ahrefs logged 9 searches in 12 logged-out tests, with only simple definitions answered from memory. ChatGPT head Adam Fry told Big Technology its web searches total about one-fifth of Google query volume, while [OpenAI](https://www.startuphub.ai/startups/openai) Signals counted 20.2% of consumer messages as information seeking in September 2026. It does not search your words. OpenAI says ChatGPT rewrites a prompt into targeted fan-out queries and may add memory or coarse location. Ahrefs saw about 1.6 fan-out queries per searched answer across 66.6 million responses from August to September 2026. About 6.8% used site: and three in four of those hit brand sites, while use of "official" rose to 13.1% in September. Citation is picky and unstable. In 4,000 US prompts from [Ahrefs Brand Radar](https://www.startuphub.ai/ai-news/market-research/2026/openai-publisher-partnerships-chatgpt-citations) in October 2026, health answers cited government, medical and academic sites 82% of the time, definitions cited dictionaries 63% with Cambridge alone at 42%, and brand questions cited company sites 59%. Cited pages matched the question more closely than skipped pages, were 458 days fresher than Google organic results in a 17 million citation study, and came from a site that ranks on Google 31.8% of the time but from the exact ranking page only 10% of the time. Glen Allsopp tracked 1,300 prompts daily for four weeks and found only about a quarter of cited pages repeated the next day. The other two feeds are narrower. Licensed publisher summaries sit alongside search for news, reviews and recipes, and you only enter through coverage in those outlets, per Ahrefs. Your own context, prompts, files, memory and connected apps, shapes answers whenever it is present, and Microsoft Bing is named as a search partner with [Shopify](https://www.startuphub.ai/startups/shopify) for shopping plus OpenAI crawler OAI-SearchBot for its own index. For search and content teams the job changes. Keep official pages fresh and crawlable by OAI-SearchBot, block GPTBot only if you want out of training, and chase the publishers OpenAI licenses. If you do not rank for the fan-out query, you do not get cited. --- Original analysis from [startuphub.ai](https://www.startuphub.ai), the #1 AI startup directory. © StartupHub.ai. All rights reserved. You may not republish this article in full without a license. Search engines and AI research tools may crawl and summarize for reference. Bulk reproduction or model training on this content requires a license. See https://www.startuphub.ai/terms.