Who Pays When the Machine Does the Reading?

It’s just after five. I’m on the sofa, the tea has long gone cold and the wine is still too warm. I’m scrolling through my LinkedIn feed when a post by Steffen Meier catches my eye. A simple question, asked with uncomfortable precision: What if tomorrow’s readers never visit our pages themselves, but simply send their AI agents instead? Steffen names the thing directly — „Model Context Protocol,“ or MCP — and frames it as the moment an IT topic becomes a strategic question for publishers and content creators. A good moment to sit with a few questions and think them through seriously, cold tea and warm wine in hand.

My first instinct was straightforward: if agents are going to curate instead of humans browsing, I just need to keep producing good, clearly structured content that’s compelling enough for a machine to pass on. SEO and GEO, taken one step further. And that’s not wrong. An agent assembling the most interesting texts on a given topic will favour what already counts in terms of mentions — or, in AI parlance, citations — from ChatGPT or Perplexity: authoritative, current, well-structured.

But that’s the smaller problem. The bigger one sits a level deeper, right where Steffen is pointing: whether the machine can even get in, technically and legally, regardless of how good my text is. This is not primarily a question of how I draw readers to my site, but of the conditions under which I receive their machines at all. MCP is explicitly an access and protocol layer — authentication, rights, licensing terms. The best article is worthless if a publisher hides content behind a paywall that simply excludes agent access. (This paragraph was written by Claude and subsequently checked with a loanword tool.)

Then comes what is known as the zero-click problem. Even if my text is the best one the agent finds and passes on, that doesn’t bring me a single visitor. The reader stays with the agent. I remain the source in the background — invisible, unpaid.

This very week, while LED walls glow at DMEXCO, champagne glasses clink and start-ups run through their pitches under artificial light, the talk is mostly about money and clicks. This genuinely strategic shift — who will even remain visible going forward — barely features in the panels.

So if zero-click is the reality, at least give it a new currency. What if you measured how often agents access a piece of content, how often they actually cite it and incorporate it into a response? That is exactly what is being built right now — though it carries the rather less poetic labels of „Pay Per Crawl“ or, more recently, „Pay Per Answer.“ Cloudflare has had a billing model for this in place since 2025, built around the long-dormant HTTP status code 402 — „Payment Required“ — and since July 2026 the model has shifted away from charging for mere access toward what actually matters: payment is triggered when your content appears in an AI-generated answer. Not the click, but the citation, is what counts.

Meanwhile, the big players — the New York Times, AP, the Financial Times — are negotiating directly with OpenAI and Google. For the long tail of publishers, for everyone who will never receive a call from a licensing department, metering layers such as TollBit or open protocols like RSL are emerging instead.

It sounds like a clean solution. It isn’t quite. Without reliable verification, there is no way to tell a payment-ready agent from a scraper that simply fakes its user agent — and billing the wrong party is worse than billing nobody at all. This new currency requires a trustworthy counting authority, and right now that role is effectively being filled by a single infrastructure company, because individual publishers — and certainly bloggers like me — cannot handle this technically on their own.

Before I think further, I need to pause and sort out what is actually being discussed here, because the comment threads under Steffen’s post already have the terminology flying around in confusion.

A chat is the most familiar form: the window where I type and an AI responds, a question-and-answer exchange that ends when I close the tab. The language model behind it is simply the computational engine generating those responses — no memory, no initiative; it reacts, it does not act. An agent is something fundamentally different: a system that equips such a model with tools, assigns it a goal and then lets it take multiple steps independently, without me directing each one. It searches, it calls up services, it decides along the way what makes sense next, and reports back only when it is done or stuck. MCP, in turn, is neither of those things. It is the socket in between — the protocol through which an agent gains access to external data, databases, and yes, my articles, without requiring a bespoke integration for every single connection.

That is precisely why Steffen’s observation lands harder than it first appears. The reader he describes is no longer sitting in a chat window typing a question about my article. They have dispatched an agent that knocks on my door via MCP, reads my text, cross-references it against ten others, and delivers only the distilled essence back to the person at the other end.

Between these different modes of use lies a sweeping developmental staircase in AI adoption — one I have climbed myself over the past two or three years without quite noticing each step as I took it. At the bottom stood the search engine, which gave me links and left me to read. Then came the chat, which answered on demand but needed each question posed individually. Then the tool-equipped model, where the AI was suddenly allowed to search, calculate or run code itself, but still waited for my next click. Now we stand at the threshold of the next stage — the agent that receives a goal and finds its own way there — and on the horizon the stage after that is already taking shape: one in which such agents no longer just talk to me, but negotiate with each other, sending a request to my website, a response to my payment system, without involving a human being even once.

For publishers of my size, that is precisely the uncomfortable insight this week delivers. Not „Is my content good enough for the machine?“ but „Who actually owns the till where the machine pays?“

It is a question that no DMEXCO stage poses with this kind of clarity, because it is uncomfortable for everyone who is currently helping to build that very till.

Which leaves the question I cannot yet answer, the one that has stayed with me since I was scrolling on the sofa: if the counting authority for this new currency turns out to be a single infrastructure company, have publishers really won a new business model — or simply found themselves a new landlord?

This English version was created with the help of AI (translated from the German original) and may differ slightly from a professional human translation.