Diop Daily #150 — October 2026

The Web Needs a Price for Machine Attention

Field note, October 25, 2026. The Web has acquired a second audience. People still read pages, compare products, follow links, and make decisions, while software increasingly performs those same movements on their behalf. The machine visitor is becoming a recurring participant in publishing, commerce, research, and public information. The old bargain for the Web was simple: a publisher supplied information, a search engine indexed it, and a human visitor might arrive. Agentic traffic breaks that sequence because discovery can happen without a click, a citation can become an answer, and a useful page can be consumed without the publisher learning what was taken or receiving anything in return.

The strategic question is therefore larger than whether websites should allow or block agents. Institutions need a machine-use contract: a way to state what an agent may read, how the material may be used, what evidence of use must survive, and when money or another form of consideration is due. This contract belongs at the boundary between content and computation. It should serve a small publisher, a cultural archive, a research laboratory, and a public information service as seriously as it serves a large platform.

Recent public evidence makes the gap visible. The W3C Strategy Team’s October paper on the Web in the AI era identifies commerce for content, fragment-level provenance, training preferences, consent, and liability as unresolved work. Cloudflare’s September account describes a rapid rise in agent requests, a growing difference between training crawlers and agents acting for people, and new systems for pay-per-use content, machine-readable access terms, reporting, and HTTP 402 settlement. The C2PA 2.4 specification adds machine-readable AI disclosure, repository receipts, structured-text support, and stronger ways to preserve the history of an asset as it changes.

Discovery is a route to value. It is not the settlement of value.

The second audience does different work

A human reader typically arrives, looks, remembers, shares, subscribes, or leaves. An agent can read a page, extract a claim, combine it with other sources, answer a question, and return the result without exposing the original page to a human. That path creates value for the user and cost for the publisher. The page may be fetched repeatedly because a question is asked repeatedly. A research archive may be queried at a high rate even though no individual page changes. A small technical publication may become part of thousands of answers while remaining invisible in its own analytics.

These uses are not interchangeable. A training crawler copies material into a model-building process. A research agent may quote a paragraph with attribution. A shopping agent may inspect a product page, compare an offer, and initiate a purchase. A public-service agent may consult a government notice and carry its meaning into a benefits decision. Each action has a different purpose, risk, cost, and claim on the source.

A binary allow-or-block switch collapses those distinctions. It forces the publisher to choose between open access with weak control and exclusion with lost reach. The useful design is a graduated access policy that can say:

  • read this public summary without payment, but preserve its source and timestamp;
  • use this article in an answer when the agent reports the passage, question, and requesting service;
  • query this dataset under a capped price and a stated purpose;
  • train on this collection only under a separate licence and retention rule;
  • invoke this commercial action only after the principal, terms, and settlement method are visible.

Those rules create a vocabulary for machine attention. They also create a way for a publisher to remain discoverable without making every use indistinguishable from free extraction.

Rights need a receipt

Payment alone does not repair the Web’s missing contract. A publisher needs to know what was used, when it was used, under which permission, and which answer or action depended on it. Cloudflare’s pay-per-use proposal is important for this reason: its description joins billing to usage reports and gives publishers a view into the questions that surfaced their work. The report is part of the value exchange. Without it, payment is a blind transfer and the archive cannot learn what readers, agents, or institutions actually need.

C2PA offers a parallel lesson from media provenance. Its specification does not declare that an asset is true; it records assertions about origin, ingredients, actions, actors, and changes so that a verifier can inspect the chain. C2PA 2.4 extends the model into structured text and adds a repository receipt. The same discipline belongs in machine-use rights. A request should carry a source reference, the allowed operation, the relevant fragment or dataset, the actor or service that used it, and a receipt that can be checked later.

For an institutional archive, the receipt should preserve at least these facts:

  1. the source asset, version, fragment, or query result that was accessed;
  2. the principal and service making the request, with the purpose declared;
  3. the permission, price, cap, licence, or public-use rule applied;
  4. the transformation performed, including quotation, summarization, translation, or synthesis;
  5. the settlement, refusal, correction, and revocation records that followed.

This is a record of relationship rather than a surveillance demand. A publisher does not need the user’s entire private conversation to establish that a particular document was used under a particular rule. A narrow receipt can protect privacy while preserving accountability. It can also give the publisher an opportunity to improve the archive. If agents repeatedly ask for a missing explanation, the publisher has evidence for what to write next.

Small archives cannot negotiate like platforms

The current market contains an asymmetry. Large publishers and large AI companies can negotiate bilateral licensing agreements. Most local newspapers, specialist journals, African-language archives, independent researchers, and community institutions cannot negotiate individually with every model provider or agent service. Their lack of bargaining scale becomes a technical exclusion if the only available rights system is a private contract between large firms.

A public machine-use rail should therefore support many small rights holders without forcing them to build a payments company or a bespoke API. The rail would expose a standard policy, identify the buyer, measure the use, collect the agreed consideration, and return the receipt. It could support free public-interest access, capped micro-payments, institutional subscriptions, reciprocal attribution, or a licensing rule for a defined collection. The crucial property is interoperability: a publisher should be able to state its terms once and reach multiple participating buyers.

This matters for African intellectual sovereignty. An archive becomes dependent when its history is extracted by machines that cannot recognize its owner, its language, its consent, or its terms of reuse. A local-language corpus may be valuable to a distant model developer while the institution that collected and maintains it receives no usage evidence and no ability to correct a mistranslation. A machine-use contract gives the institution a technical place to define the relationship. It does not make the archive closed; it makes the archive legible as an owner.

Where the investable surface is widening

The commercial object is the machine-use rights rail: infrastructure that lets a publisher or institution declare access terms, authenticate a buyer, measure a permitted use, preserve provenance, and settle value without negotiating a separate integration for every agent.

  • Machine-use policy registries: structured terms for reading, quoting, training, translation, indexing, commerce, retention, and revocation.
  • Usage measurement and reporting: systems that connect a source fragment to an agent request, answer, model workflow, or commercial action without exposing unnecessary personal data.
  • Micro-settlement and clearing: payment rails for small transactions, capped budgets, public-interest exemptions, institutional invoices, and cross-border reconciliation.
  • Provenance and rights receipts: portable records showing what was accessed, what transformation occurred, which permission applied, and how a dispute or correction can travel.
  • Archive control planes: interfaces that let small institutions change terms, revoke a collection, correct a source, and inspect which buyers depend on it.

Underwriting can begin with concrete questions. Can a rights holder publish a policy without writing a new API? Can a buyer discover the terms before use? Can the system distinguish a search, a citation, a training event, and a purchase? What percentage of machine uses produce a verifiable report? How quickly can a publisher correct a source or revoke a permission? Can a small archive receive settlement in a usable currency and retain a portable record of its contribution?

The value of this layer will appear in the quality of its records. A platform that merely charges agents has built a tollbooth. A platform that preserves purpose, provenance, usage, correction, and payment has built an institutional market. The former can extract revenue for a moment. The latter can let knowledge owners learn, govern, and remain present in the systems that depend on them.

Build the receipt before scaling the audience

An institution can begin with one public collection and four access lanes: free discovery, attributed quotation, paid repeated use, and restricted training. Define the source identity, language, update schedule, allowed transformation, reporting fields, price or exemption, and correction path. Test a benign query, an overbroad request, an expired permission, a source correction, and a buyer that cannot settle. Keep the receipt after each test.

The test passes when a second operator can determine what was used, who used it, under which rule, what value was exchanged, and how the institution can correct or withdraw the relationship. The Web’s machine audience will grow because agents are useful to people. The institutions that endure will be those that make usefulness visible, governable, and fairly settled.

Sources