ChatGPT doesn't fetch every web page the same way. According to a report from The Decoder, pages from OpenAI's publisher partners arrive inside the chatbot as pre-generated summaries rather than live, scraped text. Everything else, the ordinary open web, comes in as raw snippets, and in the underlying prompt that feeds ChatGPT's answers, those partner summaries are placed ahead of all web results.
That ordering appears to matter. An analysis of 1,000 cases by a researcher identified as Konitzny, cited by The Decoder, looked at how often ChatGPT names specific businesses or sources when it generates a ranked list, think "best laptops for students" or similar queries. Partner-sourced summaries named 70 percent of the entries sitting in positions one through three. By positions seven through ten, that figure dropped to just 23 percent.
In plain terms, the content that OpenAI has paid or struck deals to license gets a better shot at showing up clearly near the top of an answer. Whatever didn't make a partner deal is more likely to get buried or dropped the further down a list it falls.
This isn't the first sign of the gap. OpenAI told its publishing partners back in 2024 that their content would get more prominent placement inside ChatGPT, a promise made as the company signed licensing agreements with outlets ranging from news organizations to magazine publishers. A separate study out of Hamburg found that content from Axel Springer's media properties showed up far more often when people used ChatGPT's web interface than when the same questions were run through OpenAI's API, which lacks the same partner integrations.
Even favored partners don't always get it right
The arrangement isn't flawless even for the publishers it's supposed to help. The Decoder's reporting notes that partner content isn't always cited correctly, meaning the summaries can misattribute or garble material even when it's coming from an outlet OpenAI has a direct relationship with.
None of this is necessarily surprising, media companies have spent two years negotiating deals specifically to get this kind of visibility inside AI chatbots, and OpenAI has been upfront that partnerships shape some of what ChatGPT surfaces. But the scale of the gap, with partner content naming entities three times as often near the top of a list as it does lower down, raises a harder question for everyone outside those deals: when ChatGPT answers a question, is it describing the internet, or just the slice of it OpenAI has licensed?
OpenAI hasn't published its own breakdown of how partner content is weighted against the rest of the web inside ChatGPT's retrieval system. As the company keeps signing new publisher agreements, and as more of the web population gets its information from a chatbot answer rather than a search results page, that gap is likely to draw more scrutiny, from publishers left out of the deals and from researchers trying to figure out what ChatGPT is actually showing people.
ADFiled by The AI Desk
Models, assistants and the companies and chips behind them, reported from what was released and what was claimed, with the difference kept clear.
More from this desk →
Be the first to comment
Join the argument. No password, just your email or a passkey.