A publisher negotiating with an AI platform usually arrives with a strong catalogue, a trusted brand and an archive that runs decades deep. The conversation changes when someone asks how the publisher’s content was actually used: where, in which answers, under which terms, and with what evidence.
Many publishers cannot answer that question. They know their articles, chapters, databases, references and archives are valuable inputs for AI systems, but they cannot show it, and in AI content licensing that gap weighs directly on their negotiating position.
In short
Proof of AI content usage is the evidence layer that shows how publisher content is used by AI systems. It answers questions that access agreements and catalogue descriptions cannot: which content was retrieved, cited, summarized, transformed, displayed, used silently, or connected to a generated answer or workflow.
This matters because AI licensing is moving from access to evidence. A publisher may be right that its catalogue improves AI outputs, but pricing, renewal, audit, compliance and dispute resolution require records.
At Citations, we use the term usage evidence to describe rights-aware records that make AI content use visible, attributable and commercially actionable. A contract shows that content was licensed, and a usage record shows how the content was used. In the licensing market now taking shape, publishers will be asked to demonstrate the value of their content as well as describe it.
Licensing is moving from access to usage
In the first wave of AI licensing, negotiations focused on access: which catalogue could be used, for what purpose, for how long and at what price. That focus is shifting toward how the content was used inside AI workflows, including whether it was retrieved, cited, summarized or transformed, whether it was used without visible attribution, whether the correct licensed version was used, and whether the use can be audited later.
These details are becoming the commercial basis of AI licensing. Content quality sets a minimum price, and observable usage is what moves it higher. A licensing team cannot price what it cannot measure, a legal team cannot enforce what it cannot trace, and a commercial team cannot defend value it cannot connect to real usage, which is why AI usage evidence for publishers is becoming the central asset in AI content licensing.
The evidence gap in licensing negotiations
Publishers have traditionally negotiated from brand authority, editorial quality, subject expertise, catalogue depth, trusted archives and professional audiences. Those strengths still matter, but they are no longer sufficient on their own.
Saying that "our content is somewhere in the AI ecosystem" is a claim. Showing that "this specific content object was retrieved, used, cited or transformed in this answer, in this context, under these terms" is evidence. If the publisher cannot produce such a record, the platform can minimize the contribution, generalize it, or fold it into the broader argument that many inputs created the output. That argument may be true, but it does not help the publisher defend the value of its own catalogue.
The courts are sending the same signal
AI training litigation increasingly turns on evidence. In Kadrey v. Meta, the court ruled for Meta on fair use based on the record before it. The decision did not settle every future AI copyright case or mean that publishers and authors cannot win, but it carried a practical warning: a market harm argument is weaker when it is not backed by empirical evidence.
The more recent Elsevier et al. v. Meta complaint points in the same direction. Major publishers are arguing that copyrighted works were used, and they are also building a detailed record on copying, sourcing, substitution, licensing harm and market effects.
Most licensing conversations will never reach a court. The standard of persuasion is rising all the same, and evidence that matters in litigation will also carry weight at the negotiating table.
Negotiating without usage records
Many publishers are right to think their content shapes AI-generated answers. A platform, however, has little commercial reason to pay more for value it can plausibly deny, minimize or treat as generic, unless the publisher can demonstrate that value.
A publisher’s content may be authoritative, essential to answer quality and useful in reducing hallucination risk. If none of that usage is observable, the publisher negotiates from a weak position, because a contribution nobody can see is easy to underprice. The question is whether the publisher can show how its content matters.
Claim vs evidence
The distinction between the two is worth making explicit.
Question | Claim about value | Proof of AI content usage |
|---|---|---|
What does it say? | Our content matters | This content was used in this way |
What supports it? | Brand, authority, catalogue quality | Usage records, rights data, retrieval logs, attribution states |
What does it help with? | Opening a conversation | Pricing, renewal, audit, enforcement, dispute resolution |
Main weakness | Easy to minimize | Requires instrumentation |
Negotiation value | Limited | Strong |
A claim can open a deal, and evidence is what holds it up through pricing, renewal and disputes. The difference is between asking to be paid and showing why payment is due.
Search metrics no longer capture the contribution
For years, publishers measured digital value through search visibility: clicks, rankings, referrals, traffic and impressions. AI interfaces increasingly give users complete answers without sending them to the original source, so the publisher’s content may still contribute while the visible signal disappears.
Pew Research Center found that Google users are less likely to click result links when an AI summary appears, and that users very rarely click the sources cited inside AI summaries.
Web analytics can show traffic loss, and citation monitoring can show visible attribution. Neither shows the full contribution of publisher content inside AI systems. A source can influence an answer without being named, be retrieved into context and never appear in the final response, support a factual claim while another source receives the citation, or shape a summary the user never traces back. It can also be used under licensing terms the publisher cannot verify after the fact. In each case the content contributes without leaving a record the publisher can use.
Access without observability is hard to enforce
A contract may define permitted use, but if the publisher cannot observe actual use, it is difficult to prove whether the terms are respected. A platform may cite some sources while using others silently, retrieve licensed content without displaying attribution, use one version when another should apply, or generate value from publisher content without leaving a usable record.
Observability is therefore part of licensing infrastructure. It produces the evidence publishers need to govern usage, defend value and negotiate from facts. The same standard applies to usage reporting in AI licensing deals: a report is useful if it gives the publisher evidence it can use, and much less useful if it only describes activity after the fact.
What Cloudflare’s September 15 defaults show about this gap
Since September 15, 2026, Cloudflare has applied different defaults to AI crawlers. New domains joining its network, and free-tier sites that have not set their own rules, now block crawlers classified as Agent and Training on any page carrying advertising, while Search remains allowed. Existing customers kept their previous configuration unless they changed it.
Cloudflare’s data, published on July 1, 2026, shows that AI training accounted for 52% of crawler requests, up from 22% a year earlier, and that mixed-use crawlers, which combine Search and Training under one identity, made up over a third of crawler activity. Cloudflare places Google in that mixed-use category and says that, because Googlebot has not been split, it keeps access roughly twice as broad as competitors that declare separate bots for separate purposes.
This has licensing consequences. Mixed-use crawlers are governed by the strictest rule that applies to any of their behaviors, so a publisher that blocks Training to keep model builders out also blocks Googlebot’s contribution to organic visibility, unless it has changed the default.
SEO and infrastructure teams are right to look closely at that decision. It remains an access decision: it determines which crawlers get through, and it produces no record of what happened afterward, such as which page was fetched, whether it fed a specific answer, whether it was cited, or whether it counts toward a licensed volume.
A publisher that allows Search and blocks Training under these defaults has a defensible access policy, but it has not created any usage evidence it can bring to an AI licensing renewal. Access control and usage evidence answer different questions, and pricing discussions depend on the second.
What publishers need to be able to show
Strong AI licensing negotiations start from a usage record more than from a catalogue overview. Publishers need to be able to show, with evidence:
which content objects were retrieved;
which were cited;
which contributed without visible attribution;
which were summarized, transformed, displayed or used for grounding;
which version was used;
which right governed the use;
whether the usage was permitted under the applicable agreement;
whether attribution was required, and whether it was shown;
whether the usage can be audited after the fact;
whether the record can support renewal, repricing, compliance or dispute resolution.
These records sit at the center of AI-era licensing because AI changes the unit of value. In traditional publishing, value often sat at the level of the article, journal, book, database, platform or subscription. In AI workflows, it may surface as a claim, a paragraph, a citation, a generated summary, a recommendation or an answer. Publishers need a way to connect machine-level usage back to commercial value; without it, licensing terms stay abstract.
Proof of use has several layers
Proof of AI content usage is a record architecture made of several layers, each answering a different question.
Evidence layer | What it answers | Why it matters |
|---|---|---|
Asset identity | Which content object was involved? | Connects usage to catalogue value |
Rights context | Under which license or right? | Shows whether use was permitted |
Retrieval event | Was the content accessed or fetched? | Captures use before visible attribution |
Citation state | Was the source shown to the user? | Distinguishes visible from silent use |
Transformation | Was content summarized, translated or rewritten? | Supports rights and value analysis |
Workflow context | In which AI product or use case? | Separates low-value use from high-value reliance |
Audit trail | Can the event be verified later? | Supports renewal, disputes and compliance |
"Used by AI" covers several distinct events. Training, retrieval, grounding, summarization, display, citation and transformation create different value and need different records. The same distinction underlies AI training and competing products: the same content can improve an authorized product or help build a substitute for it, and only a usage record shows which happened.
Licensing norms are forming now
Open Markets Institute’s report on the AI content licensing market warns that deal structures and governance norms are forming now, often under conditions shaped by dominant platforms. Once established, such norms are hard to reverse.
If early AI content deals are built around opaque access, publishers may struggle to renegotiate later. If they are built around observable usage, publishers have a better basis for pricing, renewal, compliance and revenue sharing. Publishers that can document usage now are better placed to shape the terms others inherit; those that cannot may have to accept someone else’s definition of value.
This is also why recurring AI licensing revenue depends on evidence. A first payment can be based on access, while a renewal needs proof of use.
Provenance and usage evidence in the licensing conversation
Publisher content needs a trustworthy identity. A platform should know which source it is using, which version is authoritative, which rights apply and whether the content is authentic. Provenance covers that.
A provenance record shows where content came from. It cannot show on its own how often the content was used, which AI answers depended on it or what value that use created. For that reason, machine-readable provenance for reference publishers has to connect to usage records. Provenance protects trust in the asset, and usage evidence protects its value after use. Publishers need both, and the renewal discussion turns on the second, because it concerns what happened after access was granted.
AI reuse rights also need proof
The same issue arises when new AI reuse rights are granted. A university, enterprise, platform or institutional customer may receive permission to use licensed text inside internal AI systems. Once that right exists, the publisher still needs to know whether the content was used, how often, in which workflow, with what attribution and within which scope.
A permission does not generate a contribution signal, and a contract does not generate a usage record. That is the argument of AI reuse rights need usage signals: permission brings content into the system, and proof of use keeps the publisher in the value chain.
The objection: platforms already know the usage
AI platforms may already keep logs and know what was retrieved, displayed or used. On that view, publishers would not need their own evidence layer.
The party that holds the logs sets the terms of the discussion. If the platform controls the model, the interface, the retrieval layer, the usage definitions and the reporting, the publisher depends on someone else’s meter. Platform reporting remains useful, but publishers also need records they can access, audit, reconcile or produce independently. Without them, every renewal starts from the same asymmetry: the platform reports what happened, and the publisher asks for more detail.
Where Citations fits
Citations is built around this evidence gap. It helps publishers make AI usage of their reference catalogues observable, governed, auditable and commercially defensible, so that they can see how their content contributes across retrieval, citation, transformation and generated output, beyond whether a source appears in an answer.
Publishers should be able to demonstrate the value their content creates inside AI systems without guessing. Citations turns AI content usage into a record that licensing can rely on.
The next publisher advantage
The next advantage in AI licensing will go to publishers that can demonstrate how their catalogues are used, whatever the size of those catalogues. Content can be trusted, used and even essential, and still be commercially vulnerable if its use cannot be evidenced.
A useful first check for any publisher is whether it can currently show which parts of its catalogue influence AI-generated answers. If it cannot, that is the first gap to close, since in the AI licensing market now taking shape, use that cannot be evidenced is harder to monetize.
Frequently asked questions
What is proof of AI content usage?
Proof of AI content usage is evidence that shows how publisher content was used by an AI system. It can include records of retrieval, citation, summarization, transformation, grounding, display, rights context and auditability.
Why is access not enough in AI content licensing?
Access shows that a platform had permission to use content. It does not show what happened after access was granted. Publishers need usage records to support pricing, renewal, audit and dispute resolution.
How is proof of usage different from AI citation tracking?
AI citation tracking shows whether a source appears visibly in an AI-generated answer. Proof of usage also asks whether specific content was retrieved, used, transformed or contributed silently, even if it was not cited.
Why does evidence matter in AI copyright litigation?
Recent AI copyright cases show that courts and litigants focus heavily on factual records, market harm, licensing effects, copying and substitution. The stronger the evidence record, the stronger the legal and commercial position.
What should publishers ask before signing an AI licensing deal?
Publishers should ask whether the deal creates a usable record of what happens after access is granted: which assets are used, under which rights, in which workflows, with what attribution, and whether the record can support audit and renewal.
Continue the evidence chain
Start with the broader framework: AI Usage Evidence for Publishers
Test who audits the number: Usage Reporting in AI Licensing Deals
Prepare the renewal conversation: One-Time vs Recurring AI Licensing Revenue
Understand the foundation layer: Machine-Readable Provenance for Reference Publishers
See why permission is not a signal: AI Reuse Rights and Usage Signals
Separate authorized use from substitution: AI Training and Competing Products
Ready to assess your evidence layer? Book an AI usage evidence assessment
Sources
Kadrey v. Meta — legal analysis of the fair use ruling and the role of empirical market harm evidence
Client Alert: Summary and Strategic Analysis of Judge Chhabria’s Fair Use Ruling
Goodwin — "Northern District of California Judge Rules That Meta’s Training of AI Models Was Fair Use"
Northern District of California Judge Rules That Meta’s Training of AI Models Was Fair Use
Association of American Publishers — "Publishers and Authors File Class Action Lawsuit Against Meta and Zuckerberg"
Publishers and Authors File Class Action Lawsuit Against Meta and Zuckerberg
Elsevier et al. v. Meta Platforms, Inc. — Complaint PDF
Elsevier et al. v. Meta Platforms, Inc. Complaint
Pew Research Center — "Google users are less likely to click on links when an AI summary appears in the results"
Google users are less likely to click on links when an AI summary appears in the results
Open Markets Institute — "Same Gatekeepers, New Tollbooths: Mapping the AI Content Licensing Market"
Same Gatekeepers, New Tollbooths: Mapping the AI Content Licensing Market
Brookings — "Same gatekeepers, new tollbooths in the AI content licensing market"
Same gatekeepers, new tollbooths in the AI content licensing market
Cloudflare — "Content Independence Day, one year on: building the business model for the agentic Internet," July 1, 2026
Content Independence Day, one year on
Cloudflare — "Your site, your rules: new AI traffic options for all customers," July 1, 2026
Your site, your rules: new AI traffic options for all customers