The Fractals Agentic Media Playbook: What we know to be true in 2026

The reader stopped arriving

The funnel existed to make reader intent observable inside a space the publisher controlled — social to search to landing page to ad impression. Intent now forms inside the agent, in the chat window, in the AI summary, in the answer that never sends a click. You're optimizing for the bot that reads on someone else's behalf, not for a reader who arrives.

So every publisher serves two readers now — the human who might subscribe, and the bot that might cite. More than 50% of internet traffic is non-human, and AI-related crawling grew from 22% of crawler requests in spring 2025 to roughly half by the June 2026 peak. People get their answers inside AI surfaces, and the open web keeps a shrinking share of the time they spend looking. Publishers are already preparing for what some call "Google Zero", a world where almost no traffic arrives from search referrals.

Cloudflare's network carried 63 million HTTP requests a second at the end of 2024 and 115 million by September 2026, with peaks above 150 million. Daily requests from AI agents grew more than 1,700% in a year. The agents arrived alongside the human web rather than replacing it, and human traffic in the most heavily crawled categories — retail, computer software, IT services and financial services — still fell as much as 40% in under a year.

The search channel is closing

Zero-click searches reached 69% of Google searches in May 2025, up from 56% a year earlier, on SparkToro and Similarweb's clickstream panel, and the same panel put the US at 68% across the first four months of 2026. An independent analysis of 69 million sessions puts AI Mode's zero-click rate above 92%. Axios publisher Nicholas Johnston put it plainly in August 2026: "If your tools for audience acquisition is SEO, you're in trouble. The data is in front of you." What replaces it is citation density, whether GPT, Claude and Gemini name you when someone asks about your domain.

The measurement went with it

Today you measure opens, clicks, time on page. In agent-execution mode you get none of them, because the agent extracts what it needs and delivers the answer. No pageview, no click, no open. The one signal left is whether the agent keeps querying your source, and you only learn that if the agent provider chooses to tell you. They have no incentive to. Engagement-based advertising dies with the visit that fed it.

Agents read publishers, and that is the opening

Cloudflare's network data, drawn from more than 20% of the web, shows AI-related crawling at roughly half of all crawler requests, up from 22% in spring 2025. LLMs do not crawl the open web evenly. They pull from publisher content more than from anything else, which makes journalism the training data for AI answers. Fewer than 1% of sites block search crawlers while 17% block AI training, and in July 2026 Cloudflare split its single block switch into separate Search, Agent and Training controls on every plan. Apple, Google and Microsoft have committed to honor a no-training instruction that keeps a site indexed for search.

Whoever maps user intent to information sources owns the relationship, and OpenAI, Google and Apple are each building it — ChatGPT as a superapp, Gemini inside Workspace, context on the device. None of that is a play a publisher can make. The publisher's position is narrower and more durable: be the source an agent has to reach for when a particular class of question comes up. Cloudflare's traffic data shows why that beats publishing more, because an agent comes back every time someone asks about your subject.

Hotels reached the same conclusion early. Marriott CEO Anthony Capuano told a Bank of America investor conference in September 2026 that AI's biggest impact on his industry would fall on distribution, and Marriott responded by feeding Google and OpenAI as much content about its portfolio as it can so its hotels surface in the answers. Publishers holding the same position are blocking the crawlers instead. The reporting says they should be doing both.

Value moves up

Production got cheap. Specification did not

AI can write the article, generate the image, produce the video and format the newsletter. The cost of making content has collapsed. What it cannot do is decide what should exist, for whom, and why. The bottleneck moved from how fast you can produce to how precisely you can specify. Whoever can name the right problem is now the center of gravity, and everyone else is competing against a summary model.

Where that specification comes from is measurable, and the measurement points at the scaffolding rather than the tool. A randomized field experiment with 388 employees at a Fortune 500 retailer held AI access constant and varied only how people were told to use it: a rigid protocol that mandated joint use lowered document quality and cut output, while reframing the tool as a thought partner raised quality at the top of the distribution. The tool was identical in both arms. The failure rate tells the same story — S&P Global found the share of companies abandoning most of their AI initiatives before production more than doubled in a year, from 17% to 42%, with the average organization scrapping 46% of its proofs of concept.

Your sources and relationships cannot be benchmarked

Sarah Guo of Conviction put the shape of it in "The Untrainable": a benchmark is a thing you can measure, and anything you can measure is a thing you can train against. Work whose correctness is private to one firm never appears on a leaderboard. A general-purpose model has nothing to train against. Your sources, your relationships and the knowledge inside your newsroom are the parts a model cannot reproduce, because nothing public can score them.

The three things that decide your seat

Blocking the crawlers is the move that creates the scarcity. Restructuring is the move that creates something worth buying. A publisher doing only the first is holding a hostage nobody has agreed to pay for yet, and one doing only the second is publishing machine-readable content that anyone can take. The order matters less than doing both, and the rest of this playbook is the second half.

1. Structure it so a machine can find and cite it

Three kinds of content hold their value in an agent-mediated world.

Answer-content. Information built to answer a recurring agent question. Regulations, specifications, comparisons, how-tos, written so a machine can parse and attribute them, and kept current. Example: "What does a Recreational Pilot License cost in Singapore versus Australia?" An agent routes that query to whichever publisher has structured the answer cleanly, and right now that is a short list.

Signal-content. The delta, as it happens: what changed, and who it affects. Timestamped, and linked to whatever came before. Example: "CASA released a new medical standard this morning — here is what changed from the previous version and who is affected." Agents need the change, not a restatement of what already existed.

Context-content. Analysis an agent cannot synthesize because it depends on lived experience and judgment. The model links to it rather than summarizing it, because it knows it cannot reproduce the judgment. Example: "Here is what it actually feels like to fly this aircraft in South Australian summer turbulence — and here is what the spec sheet does not tell you."

Almost everything else — narrative features, hot takes, aggregated roundups — becomes inventory for agents to summarize and discard, which puts the writer producing it in competition with a summary model. Produce what the model cannot reproduce and you set the price.

So a publisher that wants to be cited has to publish more than articles, and a company that wants to be cited has to publish at all. Not in the old sense of a blog nobody reads. It means structured content on the domain you control, with clear attribution and a maintenance cadence, built to answer the recurring questions agents ask about that industry. The blog, the research hub and the data portal are citation infrastructure rather than marketing assets.

2. Publish it in more than one mode

Brian Morrissey of The Rebooting named this in October 2026: "Liquid publishing is here. Sometimes called headless content, the new approach to media is to be multimodal and have a business model that doesn't depend on people coming to you."

In practice, one piece of reporting becomes several artefacts. The page a human reads. A structured summary block at the top carrying who, what, when, how much and compared to what. A schema layer a machine reads without parsing the prose. A plain-text variant for anything that asks for markdown rather than HTML. A newsletter item built from the same spine. And, for a publisher with a buyer lined up, a feed.

This site does the cheap half of that already. Every page here has a markdown twin, served automatically when a request asks for it, so an agent gets the structure and a browser gets the design. One set of copy does two jobs, and neither version is a summary of the other.

Axios went further and sold the feed itself. Once reporting is structured, the structure is a product, and it can be sold without anyone visiting anything.

3. Build revenue that does not need a visit

More than 50 publisher-AI licensing agreements have been signed since 2023, and publishers are restructuring for the ones still to come. USA Today Co. told investors in August 2026 that it was reformatting its content to be machine-readable — testing approaches that include converting webpages into markdown — explicitly to attract future AI licensing deals. Google's Expert Intelligence, launched August 27, makes the same move from the reader's side: it lets you query ebooks you have bought from Google Play Books inside Gemini Notebook, and Google says it has been meeting authors and publishers about what comes next. Those deals prove content has value, though they reach neither most of the web nor most buyers.

There is a second route that skips the marketplace. Axios announced Axios Direct in September 2026: three content feeds that sell its newsroom's reporting to machine customers — market-moving news for investment firms that trade on it, a feed for companies whose internal AI models ingest it, and by 2027 a subscription that a person's agent can buy on its own. Deals are priced on assets under management and run two years, and Axios is taking them client-direct rather than through a marketplace like Snowflake's, which mirrors how it already runs: 52% to 54% of its revenue comes straight from clients. Axios made the choice worth watching. A publisher who sells wholesale takes the platform's terms, and one who sells direct writes their own.

The meters arrived this year. Cloudflare's Pay Per Use pays a publisher when its content is used, and the reporting matters as much as the payment — the publisher sees what was used, when, and what it earned. Its chief strategy officer says publishers use the company's bot controls to "create reliable scarcity for their content, and then negotiate better deals". Apple has offered publishers per-use payment for Siri news against a nine-figure budget. Google has been paying about 100 publishers for the pages that shape its AI answers, on a "demonstrated value" basis with no published method, and the spread is the warning: one publisher earns more than a million dollars a year, most earn under $1,000, which several said is less than 0.1% of their ad revenue.

That distribution is the argument. The market is real, and arriving early is worth more than arriving at all. Australia's parliament passed a News Bargaining Incentive in August that puts a 2.5% levy on platforms with more than A$250 million in local ad revenue, whether or not they carry a single news story, offset by any deals they strike. Refusing to negotiate now carries a price.

Both products ship as bets, with pricing and discovery unresolved. That is the honest state of the market in late 2026, and it is the reason the terms are still open.

What to do on Monday

Find out which level you are on

Adapted from Dan Shapiro's framework for software engineering, applied to journalism. The levels are a diagnostic rather than a ladder to climb for its own sake, and most publishers are stuck at 2.

The gap between level 2 and level 4 is a workflow design problem rather than a tool problem. The organizations that cross it rebuilt their whole production process around what AI can do. The ones that bought a ChatGPT Enterprise license and called it done are still at level 2.

Then work the checklist

Every item here comes from work we have done or watched operators do. None of it is theoretical.

• 1. Run the agent test on everything you publish Take your last three pieces. Paste each into an LLM and ask: "Based only on this content, what single action would you take on a reader's behalf?" If the LLM cannot name a specific action — something concrete it could do for the reader — you wrote a piece no agent will surface
• 2. Build a structured data layer Every piece of answer-content should have a machine-readable summary block at the top: who, what, when, where, how much, compared to what. This is data rather than the headline and dek, the shape of the response your content would return if someone queried it programmatically
• 3. Separate bot-readable from human-readable A bot reading your article wants structured facts with clear attribution, and the human reader wants narrative and voice. They need different surfaces, because one page cannot serve both at once. At minimum, every article carries an explicit structured summary that a machine can parse
• 4. Design for citation, not clickthrough The metric that matters is whether an agent cited you in its answer rather than whether someone clicked your link. That means clear attribution, timestamped claims, and the authority in your niche that makes you the default source an agent reaches for
• 5. Move up the levels Do not try to build a Dark Factory overnight. Move one level at a time. From level 2 to 3: stop editing AI output word by word. Write a style guide, a set of facts, and a specification, then approve or reject at the feature level. The breakthrough is better specs, not better prompts
• 6. Build a point of view AI can generate infinite content now, so a clear public point of view about where your market is going is the thing that makes you cite-worthy and the reason an agent picks your source over the dozens of others covering the same topic
• 7. Make your journalism verifiable, not just readable AI systems drift rather than crash, which is why their failures are hard to spot. AI-generated journalism behaves the same way. Build verification into the spec: the facts that must appear, the claims to avoid, and the one outcome the reader should be able to act on. The alternative is visible already: Trellner Research found three affiliated sites publishing 215,000 machine-generated "best software" pages, and Perplexity cites them. Whether an agent can check your work is what keeps it citing you

And hold yourself to the new numbers

Stop measuring: pageviews, sessions, time on page, bounce rate, social shares.

Start measuring:

Citation rate. How often do GPT, Claude and Gemini cite your content for queries in your domain? Run systematic probes monthly. This is your new traffic number.

Citation position. When you are cited, are you the primary source or one of many? Being first-cited in an agent's answer is worth more than being linked fifth.

Actionability score. For every piece: can an agent infer a specific action from it? Track the percentage of your output that passes this test. Target: above 70%.

Spec-to-output ratio. How much time do you spend writing specifications versus reviewing AI output? As you move up the levels, this ratio should shift toward spec-writing.

Direct relationship depth. Wholesale margins compress, and what holds them up is a direct relationship with readers who chose you. Track newsletter subscribers, paying members, and community participants.