When the Reader Is a Robot: I Tried to Fix Ads for AI Agents, Then Killed the Idea

For about twenty years the web ran on a quiet handshake. Publishers gave away content for free, search engines sent them traffic, and advertising paid the bill. Nobody signed anything, but everyone knew the terms. That handshake is coming apart, and the reason is almost too simple to state: the reader is no longer a person. It is an agent.

When an agent reads a page on your behalf and hands you the answer, you never visit the page. You never see the banner. The page still costs money to make, and now it earns nothing for the making. So I asked myself the question that the whole industry is quietly panicking about: what happens to advertising when nobody is looking?

I thought I had an answer. I spent a week building the case for it, and the same week taking it apart. This is the honest version of both.

The idea I had

The intuition was that the only way to keep a human in the loop is to pay them for their attention. So I sketched a platform that does exactly that.

An agent reading content runs into an ad I have injected into the text, say a promotion from Google. Instead of the ad dying there, unseen, it lands in an inbox inside my platform: “You got an ad from Google. Take a look.” The user opens it, views it for a few seconds, maybe scaled to how long they would normally spend, and clicks to confirm they saw it. That click pays them. The ad revenue gets split, and a slice goes to the person who actually looked.

I liked it because of the revenue split. That was the move I thought made the whole thing work: advertisers keep reaching humans, humans get paid for the attention they are giving away for free, and I sit in the middle taking a cut for closing the loop. On a napkin it holds together.

It does not survive contact with the thing it is trying to fix.

Why it does not work

Start with the direction it is pointed. The entire disruption is that the human left the page. My idea pays the human to come back and stare at ads again. That is spending money to rebuild the exact friction the whole industry is spending billions to delete. In an agent world, an advertiser does not want you to glance at a banner in an inbox after the decision is already made. They want to influence the agent’s recommendation at the moment the choice happens. That is where the money is already moving, and my design points the other way.

Then there is verification, which is fatal on its own. The one thing an advertiser buys is “a real human paid attention.” That is precisely the thing that is trivial to fake. In a world full of agents, the cheapest way to click “I saw it” and collect a two-cent bounty is to have your agent click it for you. So the model manufactures the lowest-quality, most fraud-prone attention that exists, then charges a premium for it because the revenue is sliced three ways. It is a worse product sold at a higher price.

And this graveyard already has headstones. AllAdvantage paid people to keep an ad bar open on their screen, burned through roughly $200M, and folded. Paid-email schemes went the same way. Brave, with its BAT token, is the best-funded, most serious attempt at paying users for attention, and after years it is still niche, because the reward is too small to change what people do and the friction is too high to be worth it.

There is one more reason, and it is the one I find most interesting. Injecting promotional text so it slips past an agent’s faithful service to its user is, mechanically, adversarial to the agent. It is a close cousin of prompt injection, and the labs are actively training agents to filter exactly that. So the channel is hostile by construction. I would be building a business whose core mechanism the most capable companies in the world are working to neutralize. That is not a headwind. That is a wall.

The problem is real, though, and it is big

Killing the solution did not kill the problem. If anything, looking closely made the problem look larger.

Zero-click is the headline. Roughly 68% of US Google searches ended without a click in early 2026, up from around 45% a decade ago. For news queries it is worse, climbing from about 56% to 69% in a single year. US organic web traffic fell from about 2.3 billion visits to under 1.7 billion year over year. These are not soft numbers. They are the sound of the handshake breaking in real time.

The damage to the people who make the content is concrete. Penske Media, the company behind Rolling Stone, Billboard, and Variety, sued Google alleging a drop of more than 30% in affiliate revenue. Business Insider’s organic search traffic fell about 55% between April 2022 and April 2025, a slide that fed into a 21% staff cut. Other publishers report losing somewhere between half and ninety percent of their search traffic.

And the old bargain, crawl my site, send me readers, is broken at the protocol level. Per Cloudflare’s data, OpenAI’s crawler scraped sites roughly 1,700 times for every referral it sent back. Anthropic’s ratio was around 73,000 to one. Google’s old number was about 14 to one. The take has gone up by orders of magnitude while the give has nearly vanished.

The money is moving, not vanishing

Here is the part that keeps founders awake. All that attention is still out there, and someone is still collecting on it. The only thing that changed is who.

ChatGPT reached about 900M weekly active users, more than double a year earlier, and fields something like 50M shopping queries a day. US ad spend on AI search is forecast to hit about $26B by 2029, up from well under 1% of search ad spend in 2025. Agentic commerce forecasts are all over the map, from about $144B by 2029 on the conservative end to multiple trillions on the breathless end, and the honest reading is to treat the trillions as TAM rhetoric and the low hundreds of billions as the number you would actually plan against. Meanwhile AI-referred shopping traffic grew roughly 393% year over year in early 2026, and it now converts better than non-AI traffic, a reversal from just a year before.

So the money went somewhere. Figuring out where it went is the whole game now.

What the big players are already doing

A lot, and fast. It helps to see the field in layers.

There are answer-time ad networks, the SDK middlemen that drop sponsored results into an agent’s reasoning: ZeroClick, Koah, Kontext, AgentVine, and others. There are the incumbents wiring ads straight into their own answer engines: OpenAI launched ChatGPT ads in February 2026, Google is putting ads in AI Mode, Perplexity launched then paused. There are the commerce and payment rails, which is where the heavyweights are really planting flags: OpenAI with Stripe on an Agentic Commerce Protocol, Google with its Universal Commerce Protocol and Agent Payments Protocol, Coinbase with x402, plus agent protocols from Visa and Mastercard. There is publisher-side monetization: Cloudflare’s Pay Per Crawl charging bots per fetch over HTTP 402, TollBit, ProRata. And there is a standards war on top of all of it, between the Ad Context Protocol pitched as the OpenRTB for the AI era and a competing roadmap from the IAB Tech Lab.

Notice what the incumbents are doing with their own product surfaces. They are standardizing the commerce and payment layer while keeping their answer results organic and unsponsored. That structurally limits how much pure ad injection can ever happen on the surfaces that matter most. The pure-ad path is being fenced off by the same companies that own the land.

Where the durable value actually sits

Once I stopped asking “how do I get the human to see the ad” and started asking “how does value get captured when the agent is the one transacting,” the picture clarified fast.

The durable value is not in serving the ad. It is in owning the rail and the trust layer. Three directions look real. Content-payment rails, charging agents and crawlers for access, which aligns the incentive because you only get paid when something real happens. Commerce in the loop, taking a cut of agent-driven purchases, which charges on outcomes rather than impressions and so sidesteps the fraud problem entirely. And verification or measurement, proving that a real or authorized actor actually acted, which is the unsolved problem underneath all of it and the reason signed-mandate and cryptographic-settlement approaches keep showing up in every serious protocol.

That last one is the through-line. In a medium where the engagement can be faked by the reader, the scarce thing is proof. Whoever becomes the neutral system of record for “this really happened” is holding something the impression-sellers never can.

The mirror

The closest thing to my original idea already exists. A company called AgentVine is doing roughly the SDK-middleman version of it, today, with funding. Finding them told me two true things at once.

The first is that my instinct was right. I was pointed at a real fracture in how the web gets paid, and I found it early. The radar worked.

The second is that I was already late, and the wrong shape to win. The version worth building is a rails-and-standards war, and you do not win those as a solo builder with a clever napkin sketch. You win them with a war chest, a team, and a seat at the protocol table. None of which is me, right now.

Why I am staying put

So I am not building it. Not because the problem is small. It is one of the biggest open questions in how the web gets paid for the next decade. I am not building it because the version that survives scrutiny wants a company, not an indie, and pretending otherwise is how you end up with a worse, smaller AllAdvantage.

Watching from the sidelines costs nothing, and the picture gets much clearer once the protocol race shakes out. There is no prize for being early to a war you are not equipped to fight.

The takeaway

The thing I am keeping is not the idea. It is the reminder that spotting the fracture early is a completely different skill from being the right person, at the right time, to fix it. The radar was good. The timing and the shape were wrong. Both are worth knowing about yourself, and the second one is the harder, more useful thing to learn.

So I killed a good-looking idea on purpose, and I feel fine about it. What I am walking away with is the map-reading, including the part of the map that said: not you, not now.

What I am watching

A few things will tell me whether the picture is settling or still moving: whether the incumbents open their ad networks to third-party demand or keep them closed; whether one payment protocol wins or the fragmentation drags on; whether user-trust backlash produces more Perplexity-style retreats; and whether agentic commerce volume stalls on the big assistants the way OpenAI’s Instant Checkout did when the company pulled that push back in early 2026. If those break a certain way, the door I just walked past might open into a different room.