Publishers put AI licensing into their first-quarter earnings as a revenue line that mattered. Meta has signed with People Inc., with USA Today Co. and with others. USA Today Co. said the category contributed meaningful growth.
This is genuinely good news, and it arrives against programmatic advertising that has been weak for long enough to stop being called a cycle.
The problem is the distribution. Almost every deal of consequence has been struck with a large newspaper or a wire agency, and the corpus those models trained on was not drawn from a proportionally smaller share of everybody else.
Why the money finds the mastheads
Not because the big titles negotiated better. Because a licensing market pays per counterparty, and a counterparty has to be found, verified, contracted and paid.
The transaction cost of a licensing deal is roughly fixed. Legal review, indemnities, technical delivery, an audit clause — that work costs about the same whether the archive is forty million articles or forty thousand. Against a national newspaper it is a rounding error. Against a twelve-person outlet covering a county it exceeds anything the deal could return, so the deal is not offered, and the absence gets read afterwards as a judgement about the value of the content.
It is not. It is arithmetic about overheads, and it produces a market in which the price of being licensed is being large enough to be worth the paperwork.
Music had this exact problem and fixed it
A century ago a songwriter faced the same structure. Their work was performed in thousands of venues, each use was small, none was worth a negotiation, and the aggregate was a living.
The answer was not a better contract. It was an institution: a collecting society that licenses the whole repertoire at once, monitors usage, and distributes on a formula. Nobody negotiates with a bar. The bar buys a blanket licence and the money is apportioned afterwards.
That model has real defects — the formulas favour the already-played, the monitoring has always been approximate, and the societies have needed regular antitrust supervision. It is nonetheless the only arrangement anyone has found that pays small rights-holders for high-volume, low-value-per-use consumption, which is exactly what training and retrieval are.
The News/Media Alliance's collective agreement, aimed at recurring retrieval revenue for small and mid-sized publishers, is the first serious attempt at that shape in this market. Whether it works will depend on the two things that decide every collecting society: whether usage can be measured, and whether the formula is defensible to the people at the bottom of it.
Retrieval is where the argument actually is
Training and retrieval are different transactions and the distinction is doing real work here.
Training happened, largely, before anyone was paying. It is a stock: a corpus assembled once, from which no per-article accounting is now recoverable, and the litigation over it is about compensation for something already done.
Retrieval is a flow. When an assistant answers a question by fetching a live page, that is a use, it has a timestamp, and it can be counted. Which means it can be metered and paid for on a formula — the thing collecting societies need and the thing the training argument can never have.
That is why the retrieval deals matter more than their current value suggests, and it is the strongest available answer to the problem this paper described when publishers began building for readers who never click. If the answer replaces the visit, the answer has to carry the payment the visit used to.
The local exception is not sentimental
There is a straightforward reason to care about which outlets get paid that has nothing to do with civic virtue.
A model answering a question about a county budget, a school closure or a local court case is drawing on reporting that exists in exactly one place. National titles do not cover it and wire agencies do not either. That material is simultaneously the least substitutable in the corpus and the least likely to be licensed, because its owner is too small to transact with.
This desk has written about what the loss of a local newsroom does to municipal borrowing costs and about California resolving to pay per journalist because headcount was the only thing it could count. The AI licensing market is heading for the same measurement problem from the commercial side, and it has not yet had to solve it.
What to watch
Not the announced deals, which will keep arriving and will keep going to large publishers.
Watch whether any retrieval agreement publishes its distribution formula. A blanket licence without a visible formula is not a collecting society; it is a pool with a discretionary payer, and the small publishers in it will find out what they are worth only after the money has been divided.
And watch whether any assistant starts disclosing which sources it retrieved from on a given answer. That single change would convert this from a negotiation about corpus access into a metered market — and it would make it possible, for the first time, to tell whether the county paper is being used far more than it is being paid.
The reporting of AI licensing as a notable revenue contributor in first-quarter 2026 publisher earnings amid programmatic advertising weakness; Meta's content licensing agreements with publishers including People Inc. and USA Today Co.; USA Today Co.'s statement that AI licensing contributed meaningful revenue growth; the concentration of large licensing agreements among upmarket newspapers and news agencies and the warning that local and independent providers risk exclusion; and the News/Media Alliance's collective licensing agreement intended to generate recurring retrieval-augmented generation revenue for small and mid-sized publishers are as reported by Digiday, Nieman Journalism Lab and the Reuters Institute during 2026. Deal values are not public and none is asserted here. The comparison with music performing-rights organisations is our own.
Topics mediapublishinglocal news
