ChatGPT favours licensed publishers over the open web

Axel Springer, the Associated Press, and News Corp did not sign licensing deals with OpenAI for the prestige. Per The Decoder, content from these partners now arrives inside ChatGPT through a separate pipeline, pre-structured and pre-summarized, while everyone else's content still gets pulled in raw and parsed on the fly at query time. That is not a subtle difference in formatting. It is a difference in how reliably a brand's facts, framing, and attribution survive contact with the model.

The mechanism matters more than the money

Most coverage of OpenAI's publisher deals has focused on the fees: reported sums in the tens of millions annually for outlets like News Corp. That framing misses the operational point. A licensing deal does not just buy legal cover against copyright claims. It appears to buy a structural advantage inside the retrieval layer itself. Pre-summarized content, delivered through a dedicated feed, skips the scraping, chunking, and ranking gauntlet that the rest of the web has to survive to get cited at all. The open web's content still has to be crawled, broken into passages, scored for relevance, and selected at inference time, a process that is lossy by design. Partner content, per The Decoder's reporting, arrives closer to the finished product OpenAI wants to serve.

This is the part that should unsettle any communications team that has treated "being crawlable" as the finish line of AI visibility. It was never going to be enough to be indexable. The game is shifting toward who gets a structural shortcut into the model's working context, and that shortcut is now for sale, or at least for negotiation, to a small number of outlets with enough leverage and enough lawyers to extract a deal.

A two-tier web was always implicit in these deals; now it is operational

Nobody who tracked OpenAI's publisher negotiations over the past two years should be shocked that a hierarchy exists. What is new is evidence that the hierarchy is load-bearing inside the product, not just a legal arrangement sitting off to the side. If pre-summarized partner content gets priority placement in ChatGPT's answers, then the comparison is no longer "licensed vs. unlicensed web content judged on equal footing by the retrieval system." It is "pre-digested, trusted input vs. raw material the model has to interpret under time and token constraints." Those are different games, and the second one is harder to win.

For sectors that do not have News Corp's negotiating position, this is the uncomfortable bit. Financial services firms, multilateral institutions, industrial groups, and policy bodies are not going to sign nine-figure content licensing deals with OpenAI. The UN system is not going to out-negotiate the Associated Press for pipeline priority. What these organizations can control is something smaller but not trivial: how legible their own content is to a model that has to parse it raw, under pressure, at the moment of a query. That means structured data, unambiguous attribution of facts to the organization that produced them, and content written to be extracted in self-contained chunks rather than discovered through narrative build-up. None of that buys the privileged pipeline. All of it improves the odds of surviving the ordinary one.

There is also a slower-moving implication for anyone who thought "getting cited by ChatGPT" was a single, uniform outcome to optimize for. It was never going to stay uniform once money entered the retrieval layer, because money rarely leaves systems undifferentiated. A citation sourced from a partner feed and a citation scraped from an unlicensed site may look identical in a rendered answer, but one arrived through a channel OpenAI built and trusts, and the other arrived through a channel OpenAI merely tolerates. Measuring "share of voice in ChatGPT answers" without knowing which tier produced it is measuring the wrong thing.

The practical read for comms and marketing leaders: audit not just whether your organization's content shows up in AI answers, but whether it is being pulled from a primary, structured source your organization controls, or scraped incidentally from wherever it happened to land. The former is defensible and improvable. The latter is a rounding error OpenAI's retrieval system is doing you a favor by including at all.