There is a recurring moment in the development of every information platform, and it arrives later than the founders expect. The platform is built on the assumption that content on the open web is an input — available, abundant, and requiring nothing in return beyond a link. The product works. It grows. And then the people who make the content notice that the link no longer arrives, because the platform answered the question directly and the user never left.
Search engines went through this and negotiated an uneasy settlement: publishers tolerated indexing because indexing sent traffic. Social platforms went through it and largely won, because the volume of content people produced for free swamped any leverage professional publishers had. AI answer engines are going through it now, and the settlement is genuinely unclear, because the mechanism that made the search bargain work has been removed.
The bargain that is actually broken
It is worth being exact about what changed, because the debate is often conducted in terms of copyright when the operative issue is economic.
Search indexed content in order to point at it. The publisher gave up a summary and received a visitor, and the visitor could be monetised. Whether or not the copyright position was comfortable, the trade was legible and both sides could measure it.
An answer engine ingests content in order to replace the need to visit it. The publisher gives up the substance and receives, at best, a citation that a small minority of users click. The trade is no longer legible, and the measurable part of it — referral traffic — is moving in the wrong direction for most publishers. That is the grievance, and it does not depend on any particular legal theory being correct.
This is why “we cite our sources” has not settled the argument. Citation was never the consideration in the original bargain. Traffic was.
Why revenue-sharing experiments matter more than lawsuits
Litigation will eventually produce doctrine, and doctrine matters. But litigation is slow, jurisdictionally fragmented and binary in a way that serves neither side well. The more interesting developments are the commercial ones: platforms that have started paying publishers directly, through revenue shares tied to how often a publisher’s material is used in answers.
Edgewisely’s reporting on one AI company’s attempt to build a model where the press actually gets paid describes an early version of this, and the scepticism it attracted is instructive. The objection was not that paying publishers is wrong. It was that the amounts are small relative to what display advertising once produced, and that a small payment establishes a precedent about what the content is worth.
That second objection is the serious one. In any market where a new pricing mechanism is being established, the first deals set expectations that are very hard to reset. Publishers signing early are trading immediate revenue for a weaker position in every subsequent negotiation — which is exactly the trade many of them made with social platforms a decade ago, and exactly the trade they publicly regret.
The leverage question
Publishers’ negotiating position rests on a single question: can the model produce good answers without you?
For general knowledge, the answer is plainly yes. Any given publisher’s coverage of a widely reported story is substitutable, because twenty other outlets covered it and the model saw all of them. Substitutable content has no pricing power, and no amount of collective outrage creates any.
For specialised, proprietary or timely content, the answer is less clear. Original reporting that exists nowhere else, proprietary data, deep domain expertise, and breaking information all have the property that a model cannot produce them without access. That is real leverage, and it is concentrated in a much smaller set of publishers than the ones currently negotiating.
The strategic implication for publishers is uncomfortable but straightforward: the assets that survive this transition are the ones that are hard to substitute. Aggregation, summary and commentary on widely available facts were already under pressure from the economics of digital advertising. Answer engines finish the job. Original reporting, proprietary datasets and genuine expertise get more valuable, not less, because they are the inputs that cannot be obtained any other way.
Where regulators enter
The third actor in this is competition authorities, and their involvement changes the shape of the negotiation considerably.
A platform that is dominant in one market and uses that position to acquire inputs on terms unavailable to competitors is describing a familiar antitrust fact pattern. European regulators in particular have demonstrated both the willingness and the machinery to pursue platform conduct at scale — Edgewisely’s analysis of how a multi-billion-euro ad-tech penalty could reshape the limits of platform power is a reminder that the enforcement capacity is real and the remedies can be structural rather than merely financial.
For AI companies, this creates an argument for settling with publishers that has nothing to do with copyright. A demonstrable record of paying for inputs on commercial terms is a much better position from which to face a regulator than a record of taking them and arguing about fair use. Several of the licensing deals signed in the last two years make considerably more sense read as regulatory insurance than as content acquisition.
What a durable settlement would need
Any arrangement that lasts probably has three properties, none of which current deals fully have.
Usage-linked pricing. Flat annual licences are easy to sign and impossible to defend as fair, because they bear no relationship to how much value the content actually delivers. Payment tied to measured use is harder to build and far more durable.
Auditable measurement. Publishers cannot currently verify the usage figures they are paid against. No mature licensing market operates this way for long; music and film both eventually built independent measurement, and this one will need to.
Terms available to non-giants. A regime in which only the largest publishers can negotiate leaves the long tail unpaid, which is both a competition problem and a sustainability problem for the information ecosystem the models depend on.
There is also a question about what publishers should be selling, which most have not thought about carefully. The instinct is to license archives, because archives exist and are easy to package. But archives are the least valuable thing a publisher has in this market: they are historical, largely duplicated elsewhere, and already ingested. The valuable assets are ongoing access to new material, structured data that has never been published in article form, and the ability to answer questions the archive does not cover. Publishers negotiating archive licences are selling the part of the business with the least leverage and keeping the part they have not figured out how to price.
The blocking option deserves a mention too, since it is the only source of leverage most publishers have. Refusing access is costly and only works collectively, which is why it rarely holds. But the publishers with genuinely non-substitutable content are also the ones for whom blocking is most credible, and a small coalition of those has considerably more negotiating power than a large coalition of everyone.
None of that is close to existing. What exists is a period of improvisation in which every party is establishing precedent while claiming to be doing something temporary. Publishers should negotiate accordingly, and AI companies should recognise that the cheapest deals available now are likely to be the most expensive ones later.
