Publishers, technology groups and payment firms are currently writing the rules for what AI assistants may read, buy and pay for on the web. The mesh is denser than it looks. It leaves one question open: how a single article costing a few cents actually gets paid for.
At the end of October, Google, Microsoft, OpenAI and Anthropic made a commitment many publishers had long been waiting for. The four intend to honour a uniform signal with which the operator of a website can declare that its content must not be used to train AI models. What is remarkable about this agreement is less its substance than its reach. What has been agreed on is a no. On the question that matters far more to media companies commercially, namely at what price their content may be used and how that price is paid, it says nothing.
There is no shortage of rulebooks. Since late 2024 more than a dozen standards and protocols have emerged that are meant to bring order to the way AI systems deal with content, goods and money. Their authors are publishers and internet standards bodies, commerce platforms, payment companies and the advertising industry, and the questions they answer differ accordingly. The reason for all this activity is plain: as long as search engines merely indexed texts and sent readers back to the source, a single file, robots.txt, was enough to tell machines which pages they should read and which they should not. It was never more than a request: the file locks nothing out, and only well-behaved programs comply with it. As long as both sides benefited from the flow of visitors, that hardly mattered. AI systems, however, train on texts, reproduce them in their answers and are beginning to shop on behalf of their users. That calls for arrangements between parties who do not know each other and cannot negotiate one by one.
What these arrangements achieve is best shown by an unremarkable transaction. A reader asks an AI assistant about a topic. The assistant comes across an article from a media company that sits behind the paywall, and is to reproduce it in the chat. The publisher asks 25 cents for it. The example is deliberately taken from the media industry; in retail and among data providers the same questions arise, but they carry different weight there.
Price lists without a till
It starts with the question of what a publisher permits at all. Two of the standards know only the answers yes and no. TDMRep goes back to European copyright law, which allows the automated analysis of texts as long as the rights holder has not objected in machine-readable form; the standard gives that objection a technical shape. AIPREF, a project of the internet standards body IETF, is meant to produce a signal that applies to all AI providers alike. The October commitment refers to a draft from this context.
More revealing are the two standards that go beyond prohibition. With RSL, robots.txt points to a licence file in which a publisher sets out which use it permits, in which countries and for what remuneration. In the example it would say that reproduction in an answer costs 25 cents. The organisation behind RSL wants to offer such licences in bundled form, on the model of a collecting society, and in mid-November was able to present its first licensee, Yahoo’s AI answer service. CoMP, for its part, comes from the advertising industry and works like a tender in miniature: the AI system states what it needs a piece of content for, and the publisher replies with price and terms.
What the two have in common is that they name the price without collecting it. RSL refers to other methods for payment, CoMP expressly excludes it. And like robots.txt, on which they build, all four are declarations to begin with. Anyone who disregards them is not stopped by them. Publishers thus have a price list, but neither a till nor a doorman.
An ID for machines
A condition can only be enforced if one knows who is at the door. So far, a program that calls up a website states its own name, and nothing prevents it from giving a false one. A method called Web Bot Auth remedies this by replacing the self-declaration with a cryptographic signature that can be checked against a public key. It is already in widespread use, and only with it does the request become a barrier: network operators such as Cloudflare, across whose servers a large share of web traffic runs, can let identified programs through, turn unknown ones away and charge a price for entry. Cloudflare has been trialling such an offer since summer 2025, and the cloud provider AWS followed in June 2026. Blocking and charging therefore happen where the data flows through, not in the publishers’ files. The signature, however, identifies only the machine. The media company learns which assistant is asking for the article, but not for which reader it is doing so, or whether that reader is willing to pay.
That the article can technically reach the chat, by contrast, is largely solved. The Model Context Protocol, MCP for short, has established itself as a kind of standardised connector through which providers make their content and functions available to the major assistants. This protocol knows no price either.
Built for the shopping basket
For ordering and paying, the commerce platforms and payment firms have given themselves rulebooks of their own. UCP, backed by Google and Shopify, and ACP, developed by OpenAI and the payment service provider Stripe, describe how an agent does what a customer does by hand in an online shop: search, fill the basket, order. A third protocol, AP2, adds a digital power of attorney with which a user records, in verifiable form, what the agent may buy and up to what amount.
How far practice lags behind ambition is shown by the record so far. It is true that since October it has been possible in the United States to shop at several merchants at once through UCP in Google Search. OpenAI, however, withdrew purchasing directly in ChatGPT, with which ACP had launched, after half a year, and AP2 has not yet moved beyond the specification stage. Above all, though, all three are tailored to goods, baskets and merchants. A single newspaper article does not feature in them.
The trouble with small amounts
The real bottleneck is an economic one. Anyone buying through these protocols generally pays by card, and for every card payment the payment service provider charges the seller a fixed amount on top of a percentage. In Germany, Stripe charges 1.5 per cent plus 25 cents, its competitor Mollie 1.8 per cent plus 25 cents. On a purchase of 80 euros that is a footnote. On an article costing 25 cents the fee consumes the entire proceeds. It is the old problem of very small amounts, on which many a business model on the internet has foundered since the 1990s.
Two more recent protocols take it on. They draw on a code that the rules of the web have provided since 1997 and that went unused for almost three decades: a server can answer a request with the message 402, “Payment Required”. With x402 it names the price in this way, whereupon the buyer’s software pays and makes the same request a second time. Stripe’s Machine Payments Protocol works similarly and can also settle many requests in a bundle.
That the crypto exchange Coinbase is behind x402 explains the choice of means of payment. Settlement is predominantly in stablecoins, that is, digital coins whose value is tied to the dollar; Stripe admits card payments in its protocol only from 50 cents. Technically, billing per request is thus possible. The reader of a German newspaper article would, however, have to keep a wallet with digital dollars for it, which one will hardly want to ask of them for the time being.
Monitoring by self-reporting
That leaves the question of how a publisher finds out what actually happens to its content. This has been taken up by SPUR, a coalition founded by British organisations such as the BBC and the Financial Times, which now counts more than eighty media companies and associations, among them, from the German-speaking world, the BDZV, the Funke media group and the NZZ. Its first standard defines the form in which an AI system reports that a piece of content has been retrieved, cited and displayed. The weakness lies in the design: the report comes from the party that is supposed to pay, and so far none of the major providers has committed to delivering it.
Advertising, which finances a good part of the content on the web, offers no way out either. Its new rulebooks for agents concern the buying of advertising space. That an advertisement should pay for an assistant’s retrieval of an article is not provided for anywhere so far.
A patchwork with a pattern
Anyone who lays the standards side by side will first notice that they compete with one another less than their number suggests. Most of them govern different steps and can be combined. Where two have the same task, they are likely to coexist, as card and direct debit have long done.
What stands out, though, is where the mesh breaks off. A media company can declare today that its article costs 25 cents, and the assistant can identify itself and retrieve the article. For the stretch in between, the reader’s consent, the payment, the delivery and a receipt for both sides, there is no standard that would govern this for a single piece of content in one go. That is no coincidence. Each rulebook arose where its authors do their business: with the merchants the basket, with the card companies the card, with the crypto exchange the digital coin. The publishers, whose business is the single article, have so far contributed terms and measurement methods, but no payment method.
Elsewhere the picture shifts. In retail, price and availability are in the catalogue anyway, and the purchasing protocols form the core there rather than the fringe. Those who sell data or computing power through an API, on the other hand, set their terms themselves and depend above all on connection and payment. For them the small amount is the normal case: an agent that queries weather data for a day triggers hundreds of requests at a few cents each.
For those who do not want to wait
Two things follow for providers. The standards mainly govern the relationship with machines that are strangers: crawlers one did not ask for, and assistants whose operators one does not know. When a payment method will prevail here is open, and its terms are then likely to be set above all by those who wrote it.
One’s own business need not wait for that. Anyone who wants to sell content, data or services to their customers today, and to do so wherever those customers happen to be, needs none of these standards for it, but a method that handles the individual purchase from price to receipt. NemosGo, which runs this site, offers one. The provider decides what an article, a data request or a function costs, in which countries the offer applies and on what terms. Selling works the same way across all channels: on the website, in the app, in the messenger, via a QR code and in the AI assistant, connected there through MCP. The buyer approves the amount in an app or grants the agent a budget in advance. Payment is made in euros from a prepaid balance, so that even 25 cents pay off, and both sides receive a receipt. Delivery takes place only after this approval. What is offered this way is therefore not to be had without payment, by a crawler no more than by anyone else. The request in robots.txt becomes a barrier that the provider operates itself.
Six questions about a purchase by an agent
| THE QUESTION | STANDARDS IN PROGRESS | WITH NEMOSGO TODAY |
|---|---|---|
| The questionTerms. What are the terms of purchase for products, data and services, and where can I retrieve them? | Standards in progressRSL, CoMP, AIPREF, TDMRep | With NemosGo today✓ Price and terms retrievable |
| The questionConnection. How does the agent reach the offer? | Standards in progressMCP, A2A | With NemosGo today✓ via MCP and API |
| The questionAuthentication. Who is behind the agent, and may it buy? | Standards in progressWeb Bot Auth, TAP, AP2 | With NemosGo today✓ Approval and budget per agent |
| The questionPurchase process. How does a valid contract of sale come about? | Standards in progressUCP, ACP | With NemosGo today✓ Contract of sale with invoice |
| The questionPayment. How does the agent pay, even a few cents? | Standards in progressx402, Machine Payments Protocol | With NemosGo today✓ Prepaid balance, from 0.01 cents |
| The questionFulfilment. How do delivery, receipt and cancellation work? | Standards in progressno common standard; SPUR for measurement | With NemosGo today✓ Delivery, receipt and cancellation |
Open questions: Good bots or bad bots? · Geopolitics · Technology · Intermediaries · Liability · Data protection · Regulation
Those who wait for finished standards will wait a long time and may end up with a great many parties in the sales process. With NemosGo the provider decides: it sets price and terms.
Chart: NemosGo. Standards as of November 2026, a selection.
What matters here is less the technology than the question of who is in charge. With the platforms’ purchasing protocols it is the platform that determines how purchases are made, and with payment at the network edge it is the network operator. Here, price, terms, brand and customer relationship remain with the provider. The provider keeps sovereignty over its offer, whichever way the customer finds it.
One limit should be stated: NemosGo does not support the purchasing protocols UCP, ACP and AP2. An agent that speaks only these will not find its way around NemosGo.
Sources
As of 16 November 2026.
| SOURCE | DATE |
|---|---|
| Bing, commitment on the signal against trainingblogs.bing.com/search/2026/10/your-content-your-choice-clearer-preferences-for-training-and-grounding | 30 Oct 2026 |
| IETF, AI Preferences working group (AIPREF)datatracker.ietf.org/wg/aipref/about | 7 Jan 2025 |
| IETF, draft for a signal against trainingdatatracker.ietf.org/doc/draft-kelley-aipref-training-preference-signal | 30 Oct 2026 |
| W3C, TDM Reservation Protocol Community Groupw3.org/community/tdmrep | since 2021 |
| RSL, specification 1.0rslstandard.org/rsl | 10 Dec 2025 |
| Internet Licensing Collective, launch with Yahoo Scoutinternetcollective.org/press/ilc-launch-2026 | 12 Nov 2026 |
| IAB Tech Lab, CoMP 1.0iabtechlab.com/comp | 28 Apr 2026 |
| Cloudflare, Web Bot Authblog.cloudflare.com/en-us/web-bot-auth | 15 May 2025 |
| Cloudflare, Pay per Crawlblog.cloudflare.com/introducing-pay-per-crawl | July 2025 |
| AWS, Web Application Firewall with x402aws.amazon.com/about-aws/whats-new/2026/06/aws-waf-ai-traffic-monetization | 15 Jun 2026 |
| Anthropic, introduction of MCPanthropic.com/news/model-context-protocol | 25 Nov 2024 |
| UCP, specificationgithub.com/Universal-Commerce-Protocol/ucp | version of 25 Aug 2026 |
| PPC Land, universal cart in the United Statesppc.land/googles-universal-cart-reaches-us-search-and-gemini-but-not-youtube | October 2026 |
| ACP, specificationgithub.com/agentic-commerce-protocol/agentic-commerce-protocol | version of 17 Apr 2026 |
| Retail Gazette, OpenAI changes course on checkoutretailgazette.co.uk/blog/2026/03/openai-pivots-chatgpt-shopping-strategy-after-instant-checkout-underwhelms | March 2026 |
| AP2, specification 0.2github.com/google-agentic-commerce/AP2/blob/main/docs/ap2/specification.md | no date |
| Stripe, pricing and fees for Germanystripe.com/de/pricing | read 21 Sep 2026 |
| Mollie, fees for Germanymollie.com/de/pricing | read 21 Sep 2026 |
| IETF, RFC 2068, HTTP/1.1, status code 402rfc-editor.org/rfc/rfc2068 | January 1997 |
| x402, official sitex402.org | no date |
| Stripe, Machine Payments Protocol documentationdocs.stripe.com/payments/machine/mpp | no date |
| SPUR, Content Telemetry Standard 1.0spurcoalition.org/updates/spur-content-telemetry-standard-v1-published | 2 Oct 2026 |
| SPUR, The New York Times, AFP, The Economist and The Atlantic joinspurcoalition.org/updates/the-new-york-times-agence-france-presse-the-economist-and-the-atlantic-join-the-spur-coalition | 13 Oct 2026 |
Let us talk
Working on the same question, or see it differently? Write to us.