For more than twenty years, publishers built digital revenue by making authoritative content easy to find. Search engines rewarded that work with traffic, and traffic could be measured in clicks, referrals and rankings.
That model is weakening. A growing share of searches ends without a click, and AI-generated answers now give users a complete response inside a search engine, an assistant, an enterprise workflow or an agent. The publisher’s content may still supply the answer and the authority behind it, but the visit that used to record that contribution no longer happens.
A contribution that leaves no trace is hard to price. AI usage evidence is the record that restores part of that trace, and it is becoming a basis of AI content licensing.
In short
AI usage evidence for publishers is the record of how AI systems use publisher content. It links content, rights, attribution, provenance and usage events in a form that can support licensing, audit, renewal, pricing and contributor compensation.
An access agreement says that content may be used, and a platform report may say that it was used. Usage evidence shows which content was used, in which AI workflow, under which rights, with what attribution, and where the record is kept.
A usage record only covers the events it observes. A service that sits between an assistant and a catalogue can record each access that passes through it. It cannot see a copy made elsewhere, a model trained on scraped text, or the exact weight a passage carried inside an answer. The scope of any evidence layer has to be stated with it.
Why missing usage records weaken a publisher’s position
Suppose an AI system relies on your catalogue to answer thousands of questions, but you cannot show which content was used, which version, how often or under which rights. The loss of traffic is the visible symptom. The larger problem is that you have nothing to put on the table when the price is discussed.
In the next licensing cycle, a publisher’s authority will open discussions, and the terms will depend on what it can show about how its content was used in AI-generated answers. The shift from access terms to usage records is traced in proof of AI content usage.
Fluent answers and authoritative sources
A model can produce a fluent answer that is not authoritative. Authority comes from source quality, provenance, version control, rights clearance, attribution and the ability to check how an answer was built.
This matters most in STM, legal, medical, academic, regulatory, financial and professional information, where customers pay for confidence in the source as much as for the text. They need to know whether the source was authoritative, whether the right version was used, whether the content was licensed for that use, and whether the citation and the answer can be verified.
Without those controls, an AI system can sound right while resting on weak commercial, legal or scholarly ground. Provenance is the first condition, which is why machine-readable provenance for reference publishers sits under any usage record. A system that cannot identify the authoritative source cannot reliably preserve rights or attribution.
Platform reporting and records the publisher holds
Usage-based compensation depends on who measures the usage. A platform report of how often your content was used can be useful, yet it is produced by the counterparty in the negotiation.
Relying only on that report means accepting its figures without verification. That may be acceptable when little money is at stake. It is harder to accept when usage data determines revenue, rights enforcement and renewal terms.
Publishers therefore need records they hold or can audit, tied to the rights that apply, at the points where their content is served. The test for usage reporting in AI licensing deals is whether the report gives the publisher evidence it can use, or leaves it dependent on someone else’s meter.
What AI usage evidence answers in practice
At a business level, AI usage evidence answers four questions:
which content was served to which AI workflow;
which version of that content was used;
how often that content was used;
under which rights and commercial terms the use took place.
A publisher that cannot answer these questions from its own records usually negotiates on the platform’s figures. If a catalogue is used but the use cannot be shown, the publisher is commercially exposed. If content is cited but the citation cannot be verified, its authority is weaker. If rights exist in contracts but not as machine-readable signals, enforcement is manual and comes after the fact.
Usage evidence records that a passage was served for a given request. Whether that passage shaped the generated answer is a separate question, and a record of access does not settle it on its own.
Access, attribution, provenance and usage evidence
Each of these terms answers a different question, and none of them replaces the others. The table sets out what each one answers and where it stops.
Concept | What it answers | Why it is not enough alone |
|---|---|---|
Access | Was the AI system allowed to use the content? | Permission does not prove actual use |
Attribution | Was the publisher or source named? | Visibility does not capture silent use |
Provenance | Where did the content come from, and what is its status? | Origin does not prove downstream value |
Platform reporting | What does the platform say happened? | The counterparty controls the meter |
Usage evidence | What content was used, under which right, in which workflow, with what record? | Covers only the events the recording service observes |
A licensing strategy should treat these as separate requirements. Access is what gets a deal signed; usage evidence is what supports pricing, audit and renewal within its scope.
Grounding and what a publisher can later prove
Grounding is usually discussed as a technical feature, but it also determines what a publisher can later prove. A system can retrieve content without showing that it contributed, cite a source without showing how it shaped the answer, or access licensed material without leaving a usable record for the publisher.
Authoritative publishers therefore need observable access, traceable provenance, version-level records, machine-readable rights and verifiable citations. For the same reason, AI content attribution for publishers has to go beyond being named in an answer and produce a record the publisher can use.
From traffic metrics to usage records
In search, value was measured by the click. In AI systems, value depends on use: whether content was served, how, in which workflow, and under which rights.
Some platforms are moving in this direction. Microsoft’s Publisher Content Marketplace presents premium publisher content as something AI builders can license, with publishers defining usage terms and receiving reporting on how their content creates value in AI environments. The open question for publishers is whether such records can become auditable, rights-aware and held on the publisher’s side, rather than remaining platform reporting.
Retrieval, citation and clicks are separate events, and a source can appear in one without the others. That is the commercial point of retrieved is not cited: a visible citation is one surface signal, and publishers need records of the use behind the answer.
Clicks are already declining
SparkToro and Datos reported in their 2024 zero-click study that 59.7% of Google searches in the European Union and 58.5% in the United States ended without a click. Ahrefs found that Google AI Overviews reduced click-through rates to the top-ranking page by 34.5% in its April 2025 analysis.
These figures do not mean that publisher content has stopped creating value. They mean that the old measurement system captures less of it. A user may get the answer without visiting the publisher, an assistant may summarize a source without sending a referral, and an enterprise workflow may draw on licensed content without triggering an analytics event.
What an AI usage record should capture
A usable evidence layer records more than a raw access log. At minimum, each record should capture:
which content object was involved;
which version was used;
which right or license applied;
which AI workflow triggered the use;
whether the content was retrieved, cited, summarized, displayed, translated, transformed or used for grounding;
whether attribution was required, and whether it occurred;
whether the use stayed inside the licensed scope, or was refused because it did not;
whether the record can be audited later;
whether the event can support pricing, renewal, remuneration or dispute resolution.
A contract defines what is permitted, and the usage record shows what happened at the points it covers. Without that record, pricing and renewal discussions rely on estimates.
Why provenance and usage evidence need each other
Provenance establishes origin: that a content object is authentic, authoritative, current and governed by known rights. Usage evidence covers what happened after the content entered an AI workflow, such as whether it was retrieved, cited, transformed or served for a generated answer.
The distinction is set out in content provenance vs usage evidence. AI licensing needs both, because a usage record without provenance cannot say what was used, and provenance without usage records cannot say how often.
One-time deals and renewal evidence
The first wave of AI licensing has produced real revenue. A one-time deal shows that a catalogue has value, but it does not show that the value can be charged again.
To defend a renewal, a publisher needs to show which assets were used, how often, in which workflows, under which rights and with what attribution. This is why one-time vs recurring AI licensing revenue is a measurement question as much as a pricing one.
Collective management will also depend on usage evidence
AI content licensing will not happen only one publisher at a time. Collective management organizations, reproduction rights organizations (RROs) and licensing intermediaries will help turn AI revenue into remuneration, and they face the same problem at scale.
A distribution key is only as credible as its input data. Rightsholders will challenge estimated inputs, and platform-declared inputs will raise questions of independence. Inputs that count only citations miss silent use, and inputs that ignore retrieval and grounding may miss high-value contributions. What AI copyright collective management distribution keys lack is measured use as an input.
The objection: AI usage is too complex to measure
AI usage is complex. Models retrieve, rank, summarize, transform and generate; some usage happens at training, some at inference, some inside enterprise workflows and some in user-facing answers.
Publishers do not need a complete map of every internal model operation to improve their position. They need reliable records at the points where content passes through governed access, retrieval, rights checks, attribution and outputs, and a clear statement of what those records do not cover. The test is whether the records support pricing, audit and renewal for the uses they describe.
The question to ask before signing the next AI deal
Before any AI licensing negotiation, a publisher should check whether it can show which parts of its catalogue were used by AI systems, through which channel, and under which commercial terms, and which uses fall outside what it can observe.
The demand for this evidence no longer comes only from AI licensing. In July 2026, the American Library Association adopted AI guidance containing a library-specific vendor review checklist: whether prompts are logged, whether that data trains a model, how long it is retained, and what audit rights the institution holds. These are usage evidence questions, now raised through procurement.
What Citations records, and what it does not
Citations is the licensing layer between a publisher’s books and the AI assistants readers already use, such as Claude and ChatGPT. When an assistant requests a passage through the Citations connector, the publisher’s territory and rights terms are checked before anything is revealed, out-of-terms requests are refused, and each access is recorded as a usage event the publisher can audit line by line. With Self-Hosted Secure Reveal™, the text is served from the publisher’s own server and the platform keeps no copy.
That record covers the activity that passes through the Citations service. It does not observe uses outside it, such as training on copies obtained elsewhere or answers produced by assistants that do not connect to the service. Product statuses are described on the Citations page for publishers.
Frequently asked questions
What is AI usage evidence for publishers?
AI usage evidence is the record of how publisher content is used by AI systems. It can include which content was retrieved, cited, summarized, transformed or displayed, under which rights, in which workflow, and with what audit trail, for the events the recording system observes.
Why are access agreements not enough?
Access agreements define permission. They say what an AI system may use. They do not show what was actually used, how often, under which rights, or with what value. Publishers need usage records to support pricing, renewal and audit.
Is platform reporting enough for publishers?
Platform reporting is useful, but it is produced by the counterparty. If the platform controls the interface, model, logs and reporting definitions, the publisher depends on someone else’s meter.
How is usage evidence different from AI citation tracking?
AI citation tracking shows whether a publisher or source appears visibly in an AI answer. Usage evidence also captures retrieval, rights context, attribution state, transformation and auditability for the accesses it records.
Why does provenance matter for AI usage evidence?
Provenance identifies the content object, source, version, rights and status. Without provenance, usage records are weak. Without usage records, provenance does not show downstream use. Publishers need both.
Can usage evidence show how much a passage influenced an AI answer?
Not on its own. A usage record shows that a passage was served for a request, under which terms. How much that passage shaped the generated answer depends on the model, and no access record measures it directly.
What should publishers ask before signing an AI licensing deal?
They should ask whether the deal creates usable records after access is granted: which assets are used, under which rights, in which workflows, with what attribution, and whether those records can support audit, pricing and renewal.
Where to go next
How licensing practice is moving from access terms toward records of use is traced in proof of AI content usage.
Who audits the number in a licensing report, and what to require from it, is covered in usage reporting in AI licensing deals.
Why a training fee paid once rarely survives the next model version is set out in one-time vs recurring AI licensing revenue.
How collective management organizations could turn measured use into a distribution key is examined in AI copyright collective management.
Sources
Microsoft Advertising — “Building Toward a Sustainable Content Economy for the Agentic Web,” February 2026
Building Toward a Sustainable Content Economy for the Agentic Web
Informa PLC — Market Update, Partnership and Data Access Agreement with Microsoft
Informa Market Update
Interactive Investor — “Confident Informa does AI data deal with Microsoft”
Confident Informa does AI data deal with Microsoft
ProRata.AI — Company website
ProRata.AI
SparkToro / Datos — “2024 Zero-Click Search Study”
2024 Zero-Click Search Study
Ahrefs — “AI Overviews Reduce Clicks by 34.5%”
AI Overviews Reduce Clicks by 34.5%
Ahrefs — “Update: AI Overviews Reduce Clicks by 58%”
Update: AI Overviews Reduce Clicks by 58%
Open Markets Institute — “Same Gatekeepers, New Tollbooths: Mapping the AI Content Licensing Market”
Same Gatekeepers, New Tollbooths: Mapping the AI Content Licensing Market
Citations — “Citations for publishers,” product statuses checked November 16, 2026
Citations for publishers
This article is operational analysis for publishing and rights teams, not legal advice.