The economic model that has supported digital publishing for the past two decades is shifting once again. Publishers have adapted to multiple waves of change.

Open access gave way to advertising, and advertising gave way to subscriptions and metered access. Each transition required publishers to adapt their strategies and their systems, but the underlying mechanics of the Web remained relatively consistent throughout.

That consistency is now breaking down.

What’s different this time is the simultaneity of the pressures.

Costs of business continue to rise while revenue growth remains uneven. Audience behaviour is fragmented, with readers arriving from a wider range of channels and showing increasingly varied levels of engagement and willingness to pay. Perhaps the most significant pressure comes from AI, where machines are consuming journalism at scale, often without any direct commercial relationship with the publisher whose work they’re ingesting.

These dynamics are not independent. They reinforce each other, and together they are beginning to challenge long-standing assumptions about how value is created and captured in digital publishing.

The numbers behind the shift

For years, the Web’s economic bargain was simple: Publishers produced content, platforms and search engines drove traffic, and that traffic translated into revenue through subscriptions or advertising.

Traffic was the bridge between content and monetisation, and that bridge held.

It’s holding less well now. Cloudflare’s analysis from January 2026 put a precise figure on what many publishers were already observing.

For every 28,400 HTML pages requested by a prominent AI crawler, only a single visit was referred back to the originating site. Combined traffic from search engines and AI crawlers grew 19% in 2025, with some individual crawlers increasing request volume by more than 300%. Over the same period, unique visitors to U.S. news sites fell 13% and pageviews fell 17%.

The journalism is still being read. However, fewer of those reads are happening on the publisher’s own platform, where subscriptions are sold and advertising is served. 

Value has not disappeared; its capture has

This distinction matters because the instinctive response to declining traffic is to treat it as a signal that content itself is worth less. We can see the evidence points in the opposite direction.

High-quality original reporting has become more valuable as the volume of information increases and the demand for verified, trusted sources grows. The AI companies whose crawlers are driving that 300% volume increase are not consuming publisher content because it’s worthless. They’re consuming it precisely because it’s the raw material their systems depend upon.

There’s a paper in Nature from 2024 on what researchers call model collapse or what happens when AI systems train primarily on AI-generated content rather than human-originated material. The quality degrades recursively. The systems get worse.

The original journalism publishers produce is a structural input to the AI ecosystem, not an incidental one. Its value has never been more legible than it is right now.

What has changed is that most publisher infrastructure was not built to capture that value. It was built for a world of human subscribers accessing content through browsers, and it’s struggling to extend that logic to a much more complex and varied set of interactions.

What legacy systems were designed for

Most access management systems in use today were built to answer one question: Should this person see this page?

They track pageviews, enforce metered limits, and present subscription offers at defined points in the reader journey, which was sufficient for a world in which the overwhelming majority of traffic came from human readers behaving in recognisably human ways. 

Publishers now interact with a much wider range of entities. This includes engaged subscribers, occasional browsers, search-indexing bots, accessibility tools, licensed data partners, unlicensed crawlers, and AI agents making requests at volumes and frequencies bearing no resemblance to human reading behaviour.

Yet legacy systems were not built to distinguish meaningfully between these different types of visitors, let alone to apply differentiated commercial logic to each.

Access decisions are typically made only after a request has already reached core infrastructure, using mechanisms more sophisticated bots can and regularly bypass. The access logic itself is tightly coupled to subscription models in ways limiting flexibility when dealing with entities that fall entirely outside a traditional subscriber relationship. 

The result is most publishers have limited visibility into who is actually accessing their content and limited ability to respond differently to different types of demand. These systems did exactly what they were designed to do; the environment they were designed for has simply moved on.

From defense to offense

Moving from defense to offense means extending the scope of the question publishers ask at the point of access.

The question is no longer only “should this person see this page?” It’s “who is this, are they human or machine, what are they entitled to, what does this interaction represent commercially, and what should happen next?”

That requires infrastructure that can identify every access request at the network edge, before content is served, and evaluate it against a full picture of identity, channel, commercial relationship, and publisher-defined rules: A paying subscriber gets access. An anonymous reader gets a registration prompt or a metered allowance. A verified licensing partner gets API access under agreed terms. An unlicensed crawler gets whatever treatment the publisher decides is appropriate, including a licensing offer, a block, or a logged record of what’s being taken.

This is the difference between a system managing subscribers and a system governing all forms of demand. The commercial logic shifts accordingly. Subscriptions, advertising, and audience development remain central, but they’re joined by usage-based licensing, tiered machine access, and the ability to participate actively in the market for AI training and retrieval data that is forming right now.

Publishers building this capability gain something beyond revenue from new channels. They gain the ability to adjust commercial rules without waiting for engineering resources and visibility into traffic that was previously invisible to them. They also gain negotiating leverage with AI companies that cannot be exercised without first knowing who is coming, what they are taking, and what it is worth.

The window is shorter than it looks

The licensing frameworks, pricing norms, and expectations about what AI access costs: All of this is being established now, in negotiations and product decisions happening across the industry. 

Markets like this tend to set quickly. The terms agreed upon early become the baseline everyone else inherits. The publishers with the infrastructure to participate in those conversations on equal terms are the ones who will shape what the market looks like rather than accept what it has already become.

The underlying value is not in question. The journalism publishers produce remains a critical input to the AI ecosystem, a point the model collapse research makes with uncomfortable clarity for anyone building large language models on a diet of synthetic content.

The question for publishers is whether their systems are capable of acting on that value, across every type of visitor and every form of demand, or whether they are watching an opportunity pass through infrastructure built for a different era.

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .