AI cloaking: the ads only AI bots can see on Time

🇫🇷 Lire en français : Cloaking IA : les pubs que seules les IA voient chez Time

Since June 2026, Time has been serving the crawlers behind ChatGPT, Claude and Perplexity a version of its website containing ads no human reader will ever see. This ad cloaking, spotted by developer Vincent Schmalbach and documented by The Register on 5 August, marks a first: a publisher selling access to what AI reads about it. The line between optimising for AI (GEO) and cheating runs exactly there. And it matters to any business that wants to exist inside AI answers, including those across the Caribbean only now discovering that those answers already talk about them.

What exactly is Time doing with AI crawlers?  

  Tip

Time detects AI crawlers by their user-agent and serves them Markdown pages loaded with ads that human visitors never see.

When ClaudeBot (Anthropic), OAI-SearchBot (OpenAI) or PerplexityBot request a page, Time returns a stripped-down Markdown version of the article, with sponsored blocks slipped inside. Ally Bank, for instance, answers questions along the lines of “which banks offer early direct deposit?”, and supplies an answer favourable to Ally Bank. A human browser gets the usual HTML, without those ads. Same URL, two different contents, depending on who is asking.

Schmalbach’s method is simple and reproducible: he requested the same page from the same machine, changing nothing but the user-agent header. As Chrome, he received 303,235 bytes of HTML. As ClaudeBot, he received 13,409 bytes of Markdown, roughly twenty times lighter, and identical for all three crawlers served.

The setup is sold with adtech firm Mobian, whose co-founder and CEO Jonah Goodhart states the ambition plainly: “When you influence ChatGPT, you’re influencing potentially all of ChatGPT.” Mark Howard, Time’s chief operating officer, confirmed to Digiday that his teams are now pitching these “agent ads”, sold at premium rates, one per page. His argument is an accountant’s one: bot traffic already exceeds human traffic on the site most days.

Time does put a “sponsored content” label at the top of each placement, with the advertiser’s name, even though no rule requires it today. But that label speaks to a machine: once the AI rewrites the answer, none of it reaches the consumer.

What fuels my suspicion lies elsewhere: Googlebot is spared. It receives the normal HTML, with no ads added. Yet Google is precisely the one player that penalises cloaking, and the one that could cost Time dearly in search. Applying a treatment to every AI crawler except that one means knowing the rule, and choosing where not to break it.

Why is this cloaking rather than GEO?  

  Tip

Serving the same content in a machine-friendly format is GEO; serving different content depending on the visitor is cloaking.

Cloaking, presenting different content to users and to search engines, is written in black and white in Google’s spam policies, consistent with what Google’s official 2026 guide recommends. One of the two official examples Google gives describes this exact mechanism: inserting text or keywords only when a search engine requests the page, with those elements staying invisible to ordinary visitors. Time transposes the practice to AI answer engines. The ad is no longer displayed next to the content: it is injected into the raw material used to compose the answers.

The scheme will stir memories among search veterans. I have worked on the web since 1996: I watched white text on white backgrounds go by, doorway pages, and already site versions that differed depending on whether the visitor was called Googlebot or not. Those techniques worked, for a while. Then Google saw through the trick, and the penalty came down, immediate and final.

Spammers live with that just fine: their sites are disposable, built to last a few days or a few weeks, long enough to cash in before starting again elsewhere. A brand, on the other hand, has to last. That is the strange part of Time’s bet: replaying a disposable-site strategy with a title founded in 1923.

Nothing says the arrangement will stay comfortable either. GPTBot and ChatGPT-User are already refused these pages with a 406 error, while OAI-SearchBot gets through: the sorting is fine-grained, and it implies permanent maintenance as crawlers evolve. Above all, AI platforms have the same reasons as Google to protect themselves from anyone polluting their sources.

One deeper question remains, which the “sponsored content” label does not settle. In France and across the European Union, an advertisement has to be identifiable as such by the consumer. Here the label does exist in the feed, but it addresses a machine. If the AI rewrites the sponsored content and hands it back as an apparently neutral recommendation, the label has nobody left to inform. No text settles this precise case yet, but the grey area is wide open.

Kimoun serves Markdown to AI too: where is the difference?  

At Kimoun, we have factored AI into the sites we build for a long time. Our own site and our clients’ sites also serve Markdown versions: every page exists at /index.md, and adding an llms.txt is systematic, sometimes backed by an llms-full.txt. So, same practice?

No, and the difference comes down to two words: identical, and requested. Those machine versions contain the same content as the pages you read, same text, same figures, same offers. Anyone can check by typing the URL into a browser. Nothing there is reserved for crawlers, nothing is hidden from humans, and that is precisely what makes verification possible.

More to the point, we do not sniff the visitor’s identity. The client is the one asking for the Markdown format, through a standard HTTP header, exactly as a browser asks for a WebP image rather than a JPEG. This is called content negotiation, and it has been a mechanism of the web since the beginning. Time, by contrast, inspects the user-agent and decides on your behalf what you are going to receive. In one case the visitor chooses, in the other the visitor gets sorted.

It is the equivalent of a printable version: an adaptation of form, not of substance. The full method (robots.txt, llms.txt and the rest) is set out in making your site citable by AI, where I already explained why a Markdown version built separately is a false good idea. Ours is not built separately: it is generated from the same source as the HTML page, at the same moment.

So the red line does not run through the format, it runs through the content. The test fits into one question I put to every project: would you be willing to show a client the version machines read? If the answer is no, it is no longer GEO.

What does this change for a business in Guadeloupe?  

If a title like Time is monetising the layer of the web that AI reads, it means that layer has become a commercial channel in its own right. The answers from ChatGPT, Perplexity or the AI Overviews now rolled out in France shape buying decisions, and some players are already paying to slip inside them.

  Note

On 24 July 2026, we queried Google’s AI overview about 189 web and digital communication agency sites in Guadeloupe, Martinique, French Guiana and Réunion. The result: 92% of queries trigger an AI overview, but the company’s official site is the number-one source in only 55% of cases, and only 31% publish an llms.txt file. These are web professionals. Our full findings.

The same survey gives a measure of the gap: with an llms.txt, a site is its own number-one source about 67% of the time, versus roughly 50% without (Kimoun findings, 24 July 2026). A business absent from that layer of the web leaves directories, platforms, and tomorrow perhaps sponsored content, to answer in its place when a customer asks an AI which company to contact in Jarry or in southern Grande-Terre.

Time is not an isolated case. A few weeks later, OpenAI signed a deal with Yelp to feed ChatGPT with business reviews: what that deal signals for local search here, and why the Caribbean is missing from it.

How do you get cited by AI without cheating?  

  Warning

If anyone promises you a guaranteed placement in ChatGPT’s or Google’s answers, walk away: nobody controls those answers, and the techniques that claim to force them expose you to platform penalties.

The Time case is really a collision between two ages of search. The old age thinks in schemes: cloaking yesterday, link buying still today. Google is explicit on that second point, filing “links bought or sold for ranking purposes” among its spam policies, unless they carry rel="sponsored". The practice remains an openly assumed revenue model for part of the industry, and the French-speaking ecosystem is particularly well stocked: Getfluence, SEMJuice, RocketLinks, Ereferer and Boosterlink are among its most established platforms.

The age now opening is more demanding. Consultant Luca Fancello calls it “reputational SEO”: do I show up, and favourably, everywhere the AI goes looking. A reputation cannot be bought one link at a time: it gets checked, cross-referenced, earned.

In practice the foundation is well known and short: quality content that answers the questions your customers actually ask, clean JSON-LD structured data, a site technically readable by crawlers, an llms.txt file, and impeccable local consistency. That is the work we carry out in our SEO and GEO support: auditing what search engines and AI read about your site, fixing what blocks them, measuring what changes.

Time has just demonstrated that the AI layer of the web can be sold. I prefer the version where it is earned. The story is not over, and I will follow it here: whether the offer holds, how Google reacts, how the AI platforms respond. Either way, that layer already exists, and ignoring it means letting others write in it on your behalf.