Are We Starting to Advertise to Robots?

So here's something that happened while we were all busy arguing about whether AI Overviews are stealing our traffic: a publisher started selling advertising space that no human will ever see. On purpose. To brands who are paying more for the privilege.

I've been sitting with this for a few days now, turning it over like a weird rock I found on the beach, and honestly? I can't decide if it's brilliant or insane. It might be both. The interesting question isn't whether it works — it's how long the window stays open before AI crawlers catch up.

Let me explain.

The Two-Faced Website

Here's what went down. A major publisher — one of the oldest names in the game — converted its pages into markdown. That's the stripped-down, text-only format that AI crawlers find easiest to digest. Nothing remarkable there. Plenty of sites are doing this. It's basically the digital equivalent of leaving cookies out for Santa, except Santa is a large language model and the cookies are structured data.

But then they did something genuinely wild. Working with an ad-tech partner, they started inserting sponsored content — formatted as FAQs, written in the advertiser's own marketing language, wrapped in schema markup — directly into those machine-readable pages. One ad per page. Sold at a premium.

Here's the kicker: when independent analysts loaded the same URLs while pretending to be an AI crawler (which requires about as much technical sophistication as changing your hat), the sponsored FAQs appeared complete with per-impression tracking tokens. Load the same page as a normal human? Nothing. Zero. Not a trace of the advertiser anywhere.

Two versions of the same webpage. The paid version is the one only a robot will ever read.

(I had to re-read that sentence three times to make sure I wasn't having some kind of episode.)

You're Not Buying Eyeballs. You're Buying the Answer.

This is where the pitch gets genuinely seductive, and I need us all to be clear-eyed about why.

The argument being made — out loud, to brands with budgets — is that influencing a human wins you one customer, but influencing a model wins you every future conversation that model has about your category. Persuade the system once, and you've shaped what it tells the next thousand people who ask.

Whether that actually works is completely unproven. But the ambition is the point. What a brand would be buying here isn't an impression served to a shopper. It's an attempt to edit what the model believes about the brand before any shopper is involved.

If that sounds like science fiction, it isn't. It's a media buy.

Think of AI visibility as running on two clocks. There's a fast clock — live content, whatever the model can retrieve the moment someone asks a question. And there's a slow clock — the trained-in knowledge the model absorbed months ago, the deep memory that determines whether your brand is even in its consideration set before a query happens.

The durable position is the slow clock, because you don't have to re-win it every time someone asks "what's the best [your category]?" Almost everything currently being sold as AI optimisation is fast-clock work. Advertising to robots is an attempt to buy fast-clock access and have it behave like slow-clock memory. Inject a favourable description at retrieval time, hope it sticks.

It's a genuinely new lever. And honestly, as a tactic? I get why brands are interested. If you've spent the last eighteen months watching your products disappear from AI-generated answers while your competitors somehow keep showing up, the idea of directly writing what the model reads about you is intoxicating.

The question isn't whether it can work. It probably can, right now. The question is how long "right now" lasts.

The Word Nobody Wants to Say

Here's the thing everyone in this space is dancing around with impressive choreography: serving one version of a page to a crawler and a different version to a human already has a name.

It's called cloaking.

Search engines spent an entire decade penalising it. Google treated it as a spam signal precisely because it breaks the fundamental assumption that the thing the machine indexes is the thing the person sees. Entire businesses got nuked for this. People lost livelihoods. There were conference talks about how not to do it. (Not the most riveting conference talks, admittedly, but they existed.)

The current experiment is essentially a bet that AI systems won't classify it the same way — that a paid FAQ visible only to a bot is legitimate optimisation rather than deception. And right now? That bet is probably correct. The rules governing how models treat this material simply don't exist yet, and in the absence of rules, the tactic works.

But here's the timing question that should be front of mind for anyone considering this: Google didn't penalise cloaking on day one either. It took years for the detection and the penalties to catch up. The pattern is always the same — tactic works, tactic scales, platforms notice, platforms crack down. The question isn't if AI crawlers develop the equivalent response. It's when. And whether you'll have extracted enough value before that happens to justify the risk of what comes after.

(It's the SEO equivalent of eating the entire cake at someone else's party. Fantastic while it lasts. Consequences arrive later.)

The Trust Problem That Eats Itself

Here's the part that determines whether this stays viable or collapses under its own success.

At small scale, this works fine. One publisher, a handful of advertisers, carefully placed FAQs on authoritative pages — the models have no reason to distrust it. The publisher's credibility is the product, and at low volume the credibility holds.

The problem — and it's the same problem every successful advertising tactic eventually hits — is what happens when it scales. Fill enough of those pages with paid descriptions written in the advertiser's own marketing voice, and the retrieval layer starts looking about as trustworthy as sponsored search results. At which point the rational response from OpenAI, Perplexity, or Google is to discount publisher markdown wholesale.

This is the trust paradox doing what it always does, just moved one floor up. Consumer advertising depletes consumer trust over time. This version depletes trust in the corpus the models read. The mechanism that makes the inventory valuable is the same mechanism that, at scale, destroys its value.

It's the advertising industry's favourite trick: finding a pristine channel, filling it with ads, and then wondering why it stopped working. We did it to email. We did it to social feeds. We did it to podcast mid-rolls. And now we're speedrunning it with the machine-readable web.

(Humanity's ability to ruin nice things remains genuinely inspirational.)

Three Things Sellers Should Actually Take From This

For those of us spending real money on visibility — whether that's Amazon PPC, Meta ads, or increasingly AI discovery — here's what this experiment actually means.

The "premium" is a mirage. The argument for charging more is that authoritative bot impressions are scarce. They're scarce today because one publisher is doing this and the format is new. Neither condition is stable. The moment a dozen publishers open machine-only ad slots — and they will, because publishers love a new revenue stream the way I love a carbohydrate — the scarcity evaporates. You're buying a window, not a channel.

The measurement is grading its own homework. Performance is assessed by putting the same FAQ questions back to the AI engines and scoring how visible and favourable the answers became. The party selling the influence is also the party grading whether the influence worked, using questions it wrote, against a system whose internals it can't see. That's not a reason to dismiss it entirely, but it is a reason to treat early results as marketing rather than proof. We've been here before. Facebook's "we measured ourselves and we're doing great" era taught us something.

The real signal is the direction of travel. Every previous shift in digital commerce has moved brands further from controlling how they're represented — from owning the shelf, to renting the search result, to hoping the model recalls you correctly. Advertising to robots is the first move that offers to hand some of that control back, for a fee, by letting brands write the machine's reference notes directly. That's genuinely interesting, and for brands with the budget to experiment and the stomach for uncertainty, there's a real argument for testing it now while the window is open. Just go in with eyes open about the shelf life.

The Bottom Line

We're watching someone try to sell billboard space inside an AI's brain. The concept is genuinely novel. It probably works right now. And for brands with experimental budget and a high tolerance for "we'll see," there's a legitimate case for trying it while the rules haven't been written yet.

But every brand that does should be asking one question: what's my exit strategy when AI crawlers start treating this the way Google treated cloaking? Because the history of search is a history of platforms catching up to tactics. The window between "this works" and "this gets you penalised" has never been shorter than it is right now, and it's getting shorter with every cycle.

The brands who'll come out ahead are the ones treating this as a short-term experiment layered on top of a long-term foundation — structured product data, earned citations, genuine authority. The billboard inside the AI's brain is a fun side bet. The boring work is still the main bet.

• Yelp licenses its reviews, photos, and business data to OpenAI, so ChatGPT surfaces Yelp content and its Request a Quote feature in local searches. The deal signals OpenAI's push to stand up an ad business.

• A now-edited OpenAI job listing for a support delivery lead reveals plans for a publisher-facing ad business, the first public hint OpenAI wants to help publishers sell inventory, not just place ads inside ChatGPT.

• ChatGPT Ads adds oCPC conversion campaigns, average daily budgets, automatic pacing, geo exclusions, AppsFlyer and Adjust integrations, and bulk API support, closing the gap with Google and Meta for performance marketers.

• Reddit shares fall 8% after the WSJ reports it may cut Google's AI access to its content. The $60M a year deal nears expiry as AI Overviews slash publisher referral traffic, forcing a renegotiation of data value.

• Google AI Overviews now appear in 43% of searches, up from 15% a year ago, as AI Mode visits climb to 279M. Similarweb data shows Google shifting from a gateway to the open web into a destination that keeps users on-platform.

• A practitioner flags three AI overreliance traps: skimming AI summaries over source material, poking at agents on your phone mid-task to feel productive, and preferring to brainstorm with a voice agent over the human in the room.

• HeyGen launches Video Podcast, turning any doc, link, or idea into a two-host video show with studio scenes, multi-cam cuts, and B-roll in minutes. It pushes AI podcasting past audio into publishable video.

🌐 Bardeen - Automate repetitive browser tasks without code.

🖼️ Leonardo AI - Image generation with custom model training for consistent characters and assets.

🪶 Ideogram - AI image generator that nails legible text and typography in designs.

📹 InVideo - Create and edit videos from a text prompt for social and marketing.

🎞️ LTX Studio - End-to-end AI filmmaking from script to storyboard to finished video.

😍It was great, all I needed

🙂Good, not great

🙄It sucked