Journalism Is Being Rebuilt for Buyers Without Browsers

Publishing now has two audiences

For most of the commercial web's history, publishers built pages for people and let machines index them. The machines found the content, ranked it, and sent readers back to the page, where subscriptions and advertising could produce revenue. That bargain is weakening.

USA Today Co. is now testing ways to format its journalism for two distinct audiences: human readers and the AI systems that retrieve, interpret, and cite published information. According to Digiday, the company is experimenting with approaches that include converting web pages into Markdown, reorganizing metadata, and examining whether modular formats such as guides perform differently from conventional articles when accessed by crawlers. Kara Chiles, the company's senior vice president of product management, frames the effort as part of a broader generative engine optimization (GEO) strategy — optimizing content for answer engines rather than the traditional search results page.

The work extends beyond discoverability. USA Today Co. is monitoring which bots reach its sites, how its reporting appears in Google AI Overviews, and whether AI platforms send any referral traffic in return. It blocks roughly 99% of self-identified AI bots and whitelists those tied to approved relationships, according to Chiles.

Chairman and chief executive Mike Reed placed the shift on the record during the company's second-quarter earnings call on August 6, 2026, telling investors that publishers now have to "create and format content for humans and for machines." He tied the work to both current AI licensing deals and future licensing opportunities.

This is an early example of a larger change in publishing architecture. An article is becoming both a reading experience and a controlled information product. The first version attracts and informs an audience. The second can be retrieved by an answer engine, supplied to an enterprise AI system, or delivered under a commercial agreement.

Search economics are forcing the issue

Traditional search created a workable exchange. Publishers allowed indexing because search results generally sent people to their sites, and those visits could generate advertising impressions, subscription conversions, registrations, and useful first-party data.

AI-generated answers can end the journey before the publisher receives a visit.

Pew Research Center analyzed 68,879 unique Google searches drawn from the browsing activity of roughly 900 U.S. adults during March 2025. When an AI summary appeared, users clicked a conventional search result in 8% of visits; when no summary appeared, they clicked in 15%. A link within the summary itself received a click in only 1% of visits where one was present.

Within this panel, the presence of an AI summary is associated with lower click-through behavior. The data does not establish that summaries caused every observed difference, or that the same magnitudes hold across other search engines, populations, and periods. What it does show is why citation alone is a weak substitute for traffic: a publisher can supply the reporting behind an answer and receive little measurable audience activity in return.

Licensing offers another way to assign value. If AI companies want current reporting, local information, archives, or reliable subject expertise, publishers can try to sell controlled access rather than hope that attribution eventually produces a click.

Permission becomes part of the product

The legal ground under that sale is still forming. In its May 2025 report on generative AI, the U.S. Copyright Office took the position that different uses during AI development and deployment "require separate consideration" — pretraining, subsequent training, retrieval-augmented generation, and search do not collapse into a single legal category. The Office did not declare that every AI use requires a license. Its position is more qualified: fair-use analysis turns on the particular use and its facts, and voluntary licensing is workable in some contexts while its feasibility remains unresolved in others. On balance, it recommended letting licensing markets continue to develop without government intervention at that stage.

Publishers are therefore building commercial systems before the law produces a universal answer. A contract can define approved uses, compensation, attribution, security requirements, and termination rights without waiting for every copyright dispute to reach final judgment.

Blocking most AI bots while improving access for selected ones can look contradictory.

Commercially, the two actions reinforce each other. A license has little practical value if the same material stays freely available through an uncontrolled route. Default blocking creates scarcity; whitelisting converts an approved relationship into access; machine-readable delivery makes that access easier for the buyer to use.

The 99% figure requires care. It was supplied by Chiles rather than independently audited, and it applies only to bots that identify themselves as AI. Crawlers do not all announce themselves reliably, and a robots.txt instruction is a request, not a technical barrier — compliance is voluntary, which is why crawler-control vendors such as Cloudflare recommend enforcement at the network layer when a site owner needs actual blocking. The emerging access model therefore requires more than a list of prohibited user agents. Publishers need traffic monitoring, enforceable delivery rules, a way to distinguish approved systems from unapproved ones, and contracts that specify which access is permitted.

Those contracts have to be precise. Training rights should not be assumed to include retrieval rights. Permission to generate citations is different from permission to reproduce substantial passages. A live news feed has an entirely different value profile. Once access becomes a product, vague definitions become revenue leakage.

Standards efforts are beginning to lower the transaction costs. Really Simple Licensing, or RSL, is an open, XML-based standard for expressing machine-readable usage, licensing, payment, and legal terms — including arrangements such as attribution, subscriptions, pay-per-crawl, and pay-per-inference. The IAB Tech Lab's Content Monetization Protocol, or CoMP, is designed to let content owners and AI systems communicate what is on offer and agree on commercial terms before crawling or use occurs; it is meant to work with direct licenses and marketplaces, not only retrieval-augmented generation. Notably, IAB Tech Lab is explicit that CoMP does not replace access controls — the framework assumes a publisher can already enforce blocking at the delivery layer. The standards, in other words, presuppose the enforcement they cannot themselves provide. Both point the same way: permission is moving closer to the infrastructure that serves the content.

What Markdown actually changes

Markdown appeals to AI systems because it strips away most of the visual and navigational material surrounding an article. Headings, paragraphs, links, and structured fields remain, in a predictable form, and the result can require far fewer tokens to process than the original HTML page.

Cloudflare's Markdown for Agents feature illustrates how such delivery works. When a client sends an Accept: text/markdown header, Cloudflare converts the page at the network edge and returns Markdown, while continuing to serve the normal site to a browser; the response even carries an estimated token count, and Cloudflare reports token reductions of up to 80%. The same mechanism has drawn immediate objections. Because the Markdown request signals to the origin server that the visitor is an agent, SEO practitioners — and Google's own John Mueller — warned that the setup makes cloaking trivial: an origin can return one thing to the header that identifies a machine and something else to a browser. The efficiency gain and the shadow-web risk are the same feature.

For a large publisher, the operational implications go further than a file format.

A machine-facing article could carry explicit fields for its publication time, correction history, geographic relevance, licensing status, and commercial classification. A guide assembled from independent modules might let an approved system retrieve the section most relevant to a user's question.

That modularity introduces risk. A single section can lose the qualifications held elsewhere in the article. A correction may reach the human page before it reaches the machine edition. A headline, summary, or extracted passage can go stale even after the canonical article has been updated.

USA Today Co. has not announced a final architecture, and Chiles has described the field as a moving target. There is no verified basis for claiming that Markdown alone will lift citation rates or licensing revenue. The commercial value will depend on the quality of the underlying structure, the access rules around it, and whether AI systems actually use the approved version.

Advertising enters the answer layer

USA Today Co. is also working on a more sensitive proposition: making branded and sponsored content visible to large language models as a service for advertisers. Chiles told Digiday the company was developing mechanisms for this during the quarter but did not disclose the planned format or name potential customers.

The logic is understandable. If consumers increasingly ask AI systems for recommendations or explanations, brands will want representation inside the information those systems retrieve. Publishers already produce sponsored articles with professional editing, subject knowledge, and distribution. A machine-readable channel could extend that business beyond the publisher's own pages.

An AI answer, though, is not ordinary advertising inventory.

The publisher does not necessarily control the final wording, placement, or surrounding context. A model can combine sponsored material with independent reporting and information from other sources, and it can omit the label that originally identified the commercial relationship.

Existing advertising principles do not dissolve at the machine boundary. Federal Trade Commission guidance holds that advertising can be deceptive when consumers are led to believe commercial material is independent content, and advises that disclosures should survive when native advertising is republished elsewhere. That expectation becomes hard to satisfy when an intermediary generates a new answer from retrieved material. A disclosure stored only in the visual design of a web page is useless to a text-processing system. Commercial status has to travel with the content as durable metadata — and the downstream platform has to preserve it in a form the user can understand.

The market has already produced a warning. In July, Time began inserting FAQ-formatted sponsored content, labeled "sponsored content" and naming the client, into the Markdown versions of its pages, with Ally Bank and the Project Management Institute among the first buyers. In August, Digiday reported that Perplexity had blocked all Markdown advertising on Time.com from influencing its agents and its user-facing results, calling the practice deceptive. Perplexity did not explain how the blocking works, and Time did not respond to a request for comment.

The important detail is that the label did not resolve the objection. Perplexity said it protects users from deceptive practices "sponsored or not"; its concern was that paid, brand-approved messaging occupied the same machine-readable page an agent might treat as factual source material — the cloaking problem in commercial form. The lesson is therefore stronger and narrower than "platforms may reject this." Durable sponsorship metadata may be necessary, but it is not sufficient. A label on an agent-readable page cannot guarantee that a generated answer will preserve the disclosure, or let the user separate reporting from client material, and a platform may reject the technique even when the publisher labels it.

Who owns the machine edition?

A parallel publishing system creates editorial responsibilities that cannot be delegated to product or advertising teams. The machine-readable version has to stay synchronized with the human article. Corrections have to propagate. Sponsored material has to remain technically distinguishable from journalism, and the separation has to survive retrieval. Licensing metadata has to record what an approved partner may do without altering the editorial meaning of the work.

The harder question is internal authority. Commercial teams may define the licensing product, but they should not decide how journalism is rewritten for favorable treatment by an answer engine. Editors need authority over the summaries and modules that represent their reporting. Legal teams need records of which version was delivered under which permissions.

Without those controls, GEO can drift into machine-facing editorial distortion. A publisher may start choosing formats according to what models quote most readily, even when those formats weaken nuance for readers. Sponsored modules may receive technical advantages unavailable to newsroom work. A correction may become invisible to a system that cached an earlier version. These are ordinary governance failures wearing new terminology, and the terminology should not obscure them.

Access is becoming inventory

USA Today Co.'s experiment is still in development. The company has not chosen a final machine-readable format, explained how sponsored material will be presented to models, or demonstrated that the work will improve licensing revenue. The direction is nevertheless clear enough to study: publisher access is being converted into managed inventory. Unknown bots meet restrictions. Approved partners receive a cleaner product. Content structure, rights information, and commercial classification move closer to the delivery layer.

For media executives, that direction reduces to a short list of questions a credible product must be able to answer. Who accessed the material? Which version did they receive? What uses were authorized? Did later corrections reach the buyer? And for anything sponsored: does the disclosure survive after the content leaves the publisher's page?

If a publisher cannot answer those questions, it does not yet have a machine-readable business. It has merely made its journalism easier to extract.