The Agentic Web Has Arrived: Publishers Rebuilt to Sell Ads to AI Agents. Would This Work?

What is the agentic web?

That second entrance is one small piece of a much larger build. The agentic web is the internet reorganised around AI agents as its primary readers, with its own protocols, its own serving formats, and its own attention economy. A human-facing page is written to be browsed, scanned, and clicked. An agent-facing page is written to be fetched, parsed, and quoted inside somebody else's answer, usually without a visit ever landing on your analytics. TIME's markdown pages are the serving-format piece of that. The ad it sold on them is an attempt at the attention-economy piece.

The academic anchor is the 17-author paper "Agentic Web: Weaving the Next Web with AI Agents" (arXiv 2507.21206, submitted 28 July 2025), whose co-authors include Dawn Song and Pieter Abbeel of UC Berkeley. It models the shift along three dimensions, intelligence, interaction and economics, and names the market that follows: the Agent Attention Economy. That term matters more than the taxonomy debates around it. When the reader is a model, attention is allocated by retrieval and citation rather than by clicks, and the thing being bought and sold is a mention inside an answer.

Three different lineages are in circulation for this term. One traces Web1 to Web2 to Web3 to agentic. One traces the Semantic Web forward. The arXiv paper traces the PC web to the mobile web to the agentic web. The most useful split for a business owner is by who does the reading. On the search web, humans browse and choose. On the social web, feeds choose for them. On the agentic web, an agent reads, compares and decides, then reports a shortlist back. Each transition moved the point of persuasion further from your homepage, and this one moves it inside a model's answer.

The protocol layer is being built by the largest platform companies, and adoption is uneven enough to be worth naming precisely.

The gap between MCP and WebMCP is the whole story of this moment. Standards get announced faster than anyone deploys them, and the file-based signals have fared worse still: llms.txt adoption in News and Media sits at 4.8% by Rankability's June 2026 count, the slowest of any major sector, and OpenAI, Google and Anthropic have all declined to support the standard, on the grounds that their extraction pipelines already parse real HTML.

Why are publishers rebuilding their sites for AI agents?

Because the traffic already flipped, and the human side of it is falling fastest for the publishers with the least bargaining power. This is the part of the agentic web that is measured rather than predicted.

The measured picture, six datasets:

• TIME sees more bot traffic than human traffic on most days, with spikes around franchise lists, according to COO Mark Howard in Digiday's 1 July 2026 report (linked in the opening). This is a first-party operational claim by a named executive about his own business.
• Cloudflare's network crossed 50% non-human traffic, with roughly 52% of AI crawling attributed to training rather than answering.
• TollBit's panel of about 400 publisher sites recorded bot-to-human visits moving from 1:200 in Q1 2025 to 1:31 by the end of 2025, alongside a 300% year-on-year rise in bot traffic and a 9.4% quarter-on-quarter fall in human page requests between April and June 2026. TollBit sells the conversion layer that publishers use to serve agents, so read these as vendor figures from a single measurer.
• Small publishers lost 60% of their search referrals over two years, against 47% for mid-sized and 22% for large publishers, per Chartbeat's analytics network of thousands of sites, reported via Axios in March 2026. ChatGPT referrals stayed under 1% of pageviews at every publisher tier.
• AI referral traffic contracted while AI usage grew. Similarweb's GenAI Visibility Index counted 267.4M US referrals from AI platforms in October 2025 falling to 226.8M in January 2026, a 15% decline, while visits to the AI platforms themselves grew 28.6% year on year. ChatGPT accounts for roughly 92% of AI referral traffic, and AI referrals total about 1% of site visits.
• Requests from agentic browsers rose 6,900% since July 2025, per bot-detection vendor HUMAN Security. Single-source, and worth watching rather than banking.

Put those together and the publisher decision becomes rational rather than faddish. Agents are consuming more, referring less, and the referral line is not coming back on its own. A publisher that treats bot traffic purely as a cost centre pays the serving bill and gets nothing. A publisher that treats it as inventory has at least a chance of being paid.

What have publishers actually built for AI agents?

A parallel web, in the literal sense: a second set of URLs, endpoints and access rules running alongside the human site, with different content formats and different terms of entry.

Two patterns are worth separating out of that table. The first: most of what publishers built serves their own readers rather than third-party agents. These are first-party products, chatbots over their own archives, built to keep the reader on the publisher's property. The second: the genuinely agent-facing builds, TIME's markdown layer and Getty's MCP server, exist mostly for two reasons that have nothing to do with visibility, namely serving cost and the ability to charge at the point of access.

Note also who is missing. The Tow Center's May 2026 analysis found publishers discussing MCP, RSL and skill files as available options while naming none who had shipped them for editorial content access. On the news side, this parallel web is still four or five companies deep.

The runnable read for a business that is not a publisher: these builds tell you which agent surfaces will carry content in your category. Track which publishers, marketplaces and associations in your own industry ship an agent endpoint, a markdown layer or an MCP server, because those are the sources your buyers' agents will read about you through, and they are the same sources you can earn a citation in.

How do publishers make money from the agentic web?

Seven business models are live or announced across the publishing industry. Two of them have produced a disclosed dollar figure. The honest summary is that the infrastructure is running well ahead of the revenue.

Read that table the way you would read a board pack. Wiley's $49M is an outcome: audited, repeatable, sitting in a public company's revenue line. A 139% return-visit lift, a 500-publication footprint and a 6,900% rise in agent requests are activity: real, measured, and not yet money. Both types belong in the analysis. Only one of them pays salaries.

The legal picture runs on the same split screen. Press Gazette's tracker counted roughly 26 lawsuits against AI companies since December 2023 running in parallel with roughly 30 licensing deals, updated 3 August 2026. CNN sued Perplexity in May 2026. News Corp sued Brave in July 2026. More than 30 US local newspaper owners sued OpenAI and Microsoft in June 2026. Folha sued OpenAI and then settled into a licensing deal in May 2026, which is the clearest statement of the industry's actual position: litigation is a pricing tactic, not a rejection of the market.

What happened when TIME sold ads to AI agents?

The first advertising format built for AI agents collided with engine policy inside two weeks. TIME, working with ad platform Mobian, began inserting sponsored content formatted as FAQs into the markdown pages that only AI crawlers receive, one ad per page, priced at a premium on the argument that agent impressions against authoritative content are scarce. Ally Bank and the Project Management Institute were among the first advertisers.

Separate that move from the chatbot builds in the table above. Ask The Post AI and Ask FT defend reader engagement on the publisher's own property, holding a person on the site for one more question. TIME is doing something commercially different: rebuilding the advertising business itself. The inventory it lost as human eyeballs left the page is being recreated as a new class of inventory, priced and sold against agent views, on pages no human will ever open. That is why the format matters more than its first two weeks suggest, and why an engine ruling on it settles more than one publisher's experiment.

On 11 August 2026, Perplexity blocked those ads from influencing its agents and its index, calling the format "deceptive". TIME did not respond to Digiday's request for comment. Mobian disputed the characterization on accuracy grounds, arguing that a model reading the page gets current, brand-verified facts at the moment it forms its answer, and remains free to weigh them or ignore them.

Mobian's CEO Jonah Goodhart had already stated the strategic case for the format plainly: "Maybe it's more important to influence the agent than even the human, because with a human you influence one person. When you influence ChatGPT, you're influencing potentially all of ChatGPT." TIME's Mark Howard kept his own hedge attached to the launch: "We don't know yet because this is brand new, and we believe that we are paving the first path forward here."

The enforcement was not an isolated dispute between one publisher and one engine. Google and Microsoft had already ruled that serving separate markdown or JSON pages with content different from the human version violates their longstanding cloaking policies. Perplexity applied the same doctrine. No engine has published a permissive policy for advertising inside agent-facing content, and the default posture across Google, Bing and Perplexity is suspicion of any differentiated agent version.

The measurement question underneath it is still open. The Media Rating Council issued interim AI principles on 8 July 2026 covering nine areas including explainability and accountability, and flagged invalid traffic as a priority for future guidance. It has not ruled on whether an agent visit counts as valid traffic at all. Until it does, every agent-ad rate card rests on an unratified premise.

What does the agentic web mean for your business?

Your customers' agents will read this layer whether or not your business appears in it. That is the practical consequence for any company found online, from a B2B software vendor to a clinic to a marketplace. The buyer who used to open ten tabs now asks one question and receives a shortlist, and the shortlist is assembled from sources the model retrieves, most of which you do not own.

Goodhart's line about influence states the asymmetry most clearly, and it holds regardless of what happens to his ad format. Persuading one human persuades one human. Being the fact a model retrieves puts you inside every answer that model gives on the subject, for as long as the retrieval holds.

That is why presence on the agentic web is earned the same way answer-engine visibility has always been earned: through the third-party sources agents retrieve, and through content on your own site that survives being quoted out of context. Our study of what predicts citation found referring domains and content freshness doing the heavy lifting, covered in detail in what drives AI citations. A separate analysis of 6,576 pages found that only 985 of them, 15%, both ranked in Google's organic top 10 and were cited by an AI engine (the pages that win everywhere). Ranking is not a proxy for being retrieved. Engines also apply different rules to the same page, which we measured separately in the AI engine rulebooks.

Treat agent visibility as a leading indicator, not a scoreboard. Citations and mentions move first, weeks before the commercial read changes, and they predict the outcome that matters: branded demand, shorter cycles, better close rates. The failure mode is reporting mention counts as if they were revenue. The opposite failure mode, ignoring the leading indicator until pipeline drops, is worse.

How should you prepare your website for the agentic web?

Most of the preparation is unglamorous: set your crawler access policy, measure agent traffic separately, and structure content so it survives retrieval. The part that needs judgment is what not to copy from publishers.

• Do not cargo-cult the markdown rebuild. Publishers serve markdown for serving cost and monetisation control. The brand-side visibility evidence for it is null. Otterly.ai's own experiment found no visibility gain (one experiment, reported by the vendor). Wesley Flippo reported a B2B SaaS client that has served sitewide Cloudflare markdown-for-agents since 15 June 2026. After more than two months: no measurable visibility change, no citations of the .md URLs, and server logs showing the markdown pages were not being fetched (LinkedIn, 18 August 2026; single client, self-reported). Both are small samples: one experiment and one client site. Pointing the other way, the htmlRAG paper (arXiv 2411.02959) finds HTML outperforms plain text for retrieval, because the tags carry structure the model uses. My view, as a practitioner running GEO programmes and as the owner of this publication: everything we have measured, including our own citation studies, points the same way. The pages that win are well-structured, semantic HTML that serves humans and bots equally, and we have yet to see a markdown mirror earn a citation such a page would not. Plenty of teams are still experimenting, and testing a markdown layer on a section of your site is a reasonable experiment. Treating it as a guaranteed move to prepare for the agentic web is not supported by the evidence so far.
• Decide your AI-bot access stance deliberately. Block everything and you disappear from the retrieval pool that answers are assembled from. Allow everything and you donate your content to model training with nothing owed in return. Your infrastructure vendor is already choosing for you: from 15 September 2026, Cloudflare will block mixed-use AI crawlers from ad-carrying pages by default unless the AI company pays. Audit what your CDN decides on your behalf before that date, and turn the default into a policy. The mechanics of crawler access, robots directives and rendering are covered in how to build a bot-friendly website.
• Measure agent traffic separately from human traffic. Server logs identify which agents fetch which URLs and how often, and GA4 can isolate AI referrals once the sources are configured correctly, which we walk through in tracking AI referral traffic in GA4. Reporting agent fetches and human sessions in one number hides both.
• Structure content so every block survives on its own. Answer engines retrieve passages, not pages. A section that only makes sense after the two above it will not be quoted. Each block should answer one question, name the brand or entity it belongs to, and carry its own numbers and dates.

The hard rule: convert, never invent. If you serve a machine-readable version of a page, it must be a conversion of the exact content a human sees at that URL. Content that exists only in the bot version is cloaking under Google's and Bing's stated policies, and TIME is the named casualty of the first enforcement.

One thing to watch rather than build: WebMCP. With near-zero deployment and no mainstream agent calling the tools as of July 2026, implementing it now buys nothing. Revisit when a major agent ships support.

Methodology and sourcing

Research compiled 18 August 2026 from primary trade coverage (Digiday, Press Gazette, Axios, MediaPost), company disclosures (Wiley's FY2026 earnings release, Cloudflare, RSL, Linux Foundation), network datasets (Cloudflare, Chartbeat via Axios, Similarweb, TollBit, HUMAN Security) and the arXiv literature. Vendor figures are labelled as such in the text, because TollBit and Mobian both sell into the market they measure. Figures that no primary source confirmed were left out rather than rounded into the argument. No client data appears in this piece.

Written by Tina Chu, founder of Novastacks. We run AI search programmes for brands and enterprises, which means we spend most weeks inside the retrieval side of this shift rather than the publishing side.

FAQ: the agentic web