As AI companies seek high-quality content to improve their models, and publishers explore new revenue streams, licensing has become a recurring theme.
Too often, the debate stops at the deal itself: who has signed, who is negotiating and what the agreement might be worth. The more important question is whether a deal supports the publisher’s wider commercial strategy.
Publishers are confronting two related shifts. Content is being used in ways over which they have little control. Once material has been incorporated into model training, there may be no practical way to remove it, with limited visibility over how it is used and little ability to audit what has been absorbed.
Alongside this, the long-standing value exchange with search platforms is weakening. Publishers make content available on the open web in return for referral traffic, but audiences are increasingly receiving answers directly through AI overviews and large language models, often without visiting the original source.
The commercial effect is becoming clearer. PPA-commissioned research found that click-through rates from AI interfaces can be as low as 0.12%, while search has historically accounted for between 30% and 80% of traffic for many publishers. It also found that automated scraping activity had risen by around 60% over 18 months, with approximately 30% operating outside robots.txt instructions.
Retrieval-based AI systems can also generate around ten times as many server requests as traditional training crawlers, adding infrastructure costs that publishers may be absorbing without compensation.
Why some publishers are more exposed than others
Yet the impact is far from uniform. B2C publishers reported significant audience disruption at more than double the rate of B2B and hybrid publishers: 67% compared with 27%. Publishers with more than 15 direct competitors also reported higher levels of disruption than those operating in less contested markets.
And publishers with a direct audience relationship via subscriptions, events, membership, community models, are demonstrably better insulated than those dependent on open discovery. The pattern is consistent: uniqueness, specialisation and direct audience relationships buy leverage; scale and broad-catalogue, search-dependent models are exposed to more risk.
Why licensing is only part of the solution
A licensing agreement can be one part of a publisher’s response, but it is not a strategy in itself. Research commissioned by the PPA, led by Magda Woods, founder of AI consultancy And The Robots, analysed 73 publicly known AI licensing agreements. It found that the typical multi-year arrangement represented less than 2.15% of annual publisher revenue, often well below the level needed to offset audience losses.
Nor is licensing an all-or-nothing decision. Different assets may justify different approaches. Broad-topic content may benefit from exposure and attribution, while premium analysis, proprietary data and subscription-only insight may warrant stronger protection and higher pricing.
Publishers must also distinguish between the rights being granted. Training rights can provide predictable revenue but are difficult to reverse once content has been incorporated into a model. Retrieval-Augmented Generation (RAG) rights can support attribution and usage-linked revenue but may also increase server demand. In either case, the commercial terms should reflect the value of the content, the proposed use and the publisher’s ability to verify compliance.
The real risk: undervaluing content
Many publishers are struggling to access negotiations at all. Bilateral negotiations often favour scale, meaning many specialist publishers find themselves reacting to market developments rather than shaping them.
For those able to negotiate, however, the greatest risk is undervaluing the assets that give them leverage. The practical starting point is to identify the content audiences value most and can’t find elsewhere: the specialist analysis, the proprietary data, the material that isn’t easily replicated.
These should be aggressively protected and priced highest, rather than treating their entire catalogue as a single, undifferentiated archive.
What publishers should do next
There isn’t a single blueprint that works for every publisher. AI exposure will vary according to business model, scale, content type and competitive intensity, and any licensing strategy should follow from that assessment.
A publisher with limited audience impact may prioritise monitoring, crawler controls and stronger direct relationships, while treating deal-making as non-urgent. A business facing greater disruption may need to pursue protection, licensing and monetisation in parallel.
Publishers should also review crawler rules, decide which content should remain open or restricted, distinguish between training and retrieval uses, and consider the route that best matches their leverage, whether bilateral agreements, collective licensing or emerging marketplaces.
The market is moving quickly, but the foundations of resilience remain consistent: trusted brands, differentiated content and direct audience relationships.
The publishers best positioned for the AI era will not necessarily be those that sign the most agreements. They will be those who understand what makes their content valuable, protect it accordingly and build sustainable commercial strategies around those strengths.
That is the central challenge AI presents to publishing: not simply how content is licensed, but whether the industry recognises, protects and properly values the journalism, expertise and trusted information on which these technologies depend.
You can view the PPA’s Licensing Framework as a member here. Magda Woods, founder of And The Robots is writing a monthly blog for the PPA. Access the most recent article here.
