Time Now Serves Chatbots a Separate Page, Ads Included

Same address at Time, two different documents. We ran the test on August 8: request the article page with a browser's headers and you get 304,745 bytes of HTML. Request the same URL with ClaudeBot's headers and you get 13,409 bytes of Markdown. Twenty-two times lighter, not a single image, and it's this version, the one no human eye will ever see, that carries the advertising.

The Question AI Overviews Left Unanswered

We covered Google's AI Overviews finally landing in France on August 2, two years after the rest of the world watched its click-through rates erode. The editorial web ran on a trade: sites let search engines index their pages, and search engines sent readers back. The AI-generated answer kept the goods and skipped the payment.

That left a question with no good answer. What does a publisher do once the answer replaces the visit? Pull out of AI surfaces and you lose the citation without getting the click back. Do nothing and you're feeding, for free, the very thing that's replacing you.

Time just proposed a third option, and it turns its back on twenty years of instinct: stop chasing the human visitor, and start treating the machine as a paying customer.

Their Answer: The Bot Becomes the Primary Audience

Time has been converting its pages to Markdown since June 2026, working with ad network Mobian. The reasoning, in one line from chief operating officer Mark Howard, reported by Digiday: on most days, Time gets more traffic from bots than from humans, with spikes whenever a Time100 list drops.

Picture a shop owner who notices more delivery vans pulling up than customers walking in. Eventually, they redesign the storefront for the vans. That's the calculation here. The crawler isn't a nuisance to tolerate anymore, it's the majority audience, and a majority audience always ends up with its own ad sales team.

What You Write When Your Reader Is a Machine

On the "Best Inventions of 2025" collection, the human-facing version weighs 1.4 megabytes. The bot version comes in at 32,000 bytes. And twelve lines in, right after the headline, before a single line of editorial copy, a block opens up: mobian-agent-ad, campaign ally-2026-q3.

That block takes up 31.5% of the document served to the machine. It's labeled "sponsored content," nobody's hiding it. What's inside is more interesting than the fact that it's there at all: a brand definition, spec tables, a four-question FAQ, then the same questions repeated as structured JSON-LD data, just to make sure the model actually reads them.

Inside those tables sits a "source" column. It carries the advertiser's name. The brand cites itself as the reference, in a format the model will treat the same way it treats any other fact.

This isn't a banner ad you scroll past. It's a press kit dropped straight into the journalist's notebook, sources already filled in. On the Business page, the same setup runs an ad for the Project Management Institute, which explains there that its certified professionals earn 16% more than everyone else. A figure supplied by the advertiser, inside its own ad.

The AI Policy Only a Robot Can Read

The best summary of the whole situation sits inside a technical file: Time's llms.txt, the document where a site spells out what AI models are allowed to do with its content.

Request it from a browser and you get a 406 error with an empty body. Request it while identifying as ClaudeBot and you get a 200 with the full text. A publisher's AI policy, unreadable to a human visitor, perfectly legible to a robot. Fittingly, it's served through the same pipe as the ads.

The content dates back to May 2025: a blanket ban on training models on Time's content, with four named exceptions: OpenAI, Perplexity, Scale, and ElevenLabs. Anthropic isn't on that list, and yet ClaudeBot is the one receiving the ad-laden Markdown. The ban covers training, not reading, so the contradiction isn't total. Mostly, it shows that a policy written fourteen months ago never caught up with what shipped in June.

One detail nobody could make up: among the sales pitches served to machines, the brand touts customer service handled by "real humans, not bots."

The Word That Comes to Mind, and Why It Doesn't Fit

Cloaking. It's a loaded word: Google defines it in its spam policies as showing different content to users and to search engines in order to manipulate rankings, penalty attached, up to full removal from results.

Except the definition doesn't hold up. Googlebot gets the same HTML page as humans do, we checked that too: no Markdown, no sponsored block, the only difference between the two versions is a consent script. The differentiation targets conversational assistants only. Neither Vincent Schmalbach, the independent developer based in Germany who first documented the setup on August 5, nor The Register use the word.

Schmalbach offers a more accurate framing: what's hidden here isn't the advertising, it's an undisclosed audience split. Ranking isn't the stake anymore. The stake is a customer to serve, and that customer doesn't have eyes.

A Bet Nobody Has Measured Yet

Here's the part most of the coverage skipped, and Digiday is the one who pulled the counter-evidence together: nothing so far shows that Markdown improves anything. The platform Promptwatch analyzed 1.6 million citations and found that it doesn't increase the odds of being cited. Google's own guidance, per Digiday, files the format under "neither helpful nor harmful." On the advertising side, the firm Oasy says it hasn't observed any effect on advertiser results yet.

Asked about measurement, Mark Howard himself says nobody knows yet, that all of this is new, and that Time is laying down the first trail. So the company is selling inventory nobody can prove works, at a price billed as premium, to an audience that doesn't click.

The Second Web Already Has Its Own Browser

None of that is slowing anything down, because the other end of the pipe is being built at the same pace. On August 7, Cloudflare shipped Kitesurf, a hosted browser built for agents: no tabs, no themes, no extensions, and in their place, priorities that only make sense to a machine, like context window size and token cost. Twelve weeks of development, start to finish.

On one side, pages written for robots. On the other, a browser that was never designed to serve a human at all. What nobody can say yet is whether the model pays off, or how many publishers will follow: there's no public data to settle it either way.

What has already changed is real, though. If you publish anything, there's a question you weren't asking six months ago: what does your page look like when a machine requests it? Most sites have no idea, because nobody had ever needed to check. Time checked, and hung out a sign.