Programmatic advertising already runs the economy everyone's bracing for. Billions of transactions a day, bids cleared in under a hundred milliseconds, money moving between parties that have never met and never will. Nobody's in the room. Nobody could be. That's been running since before most people had a smartphone.
It runs on an identity layer that doesn't work, and everyone in the business knows it.
Cookies got deprecated, then partially undeprecated, then worked around. Fingerprinting filled the gap and got regulated. Attribution windows are a negotiation, not a measurement. If you've ever sat in a meeting where the publisher's numbers and the platform's numbers disagree and neither side can prove which one is wrong, you've already met this problem. That meeting happened every week at every publisher I worked with, and it had for twenty years before I got there.
We never fixed it. We built a patch instead - verification vendors, fraud detection, viewability scoring, brand safety. An industry whose job is to stand next to the transaction and grade it after the fact.
Grading works when the thing being faked is traffic, because you can look at a request and estimate how likely a person is behind it. It stops working when the fake thing is the person.
Then it got cheap
I've spent the past few months building on Runway's video generation, close enough to it to send them notes on what breaks. Three years ago AI video meant that grainy Will Smith eating spaghetti clip everyone passed around. Now it's cinematic, and I can't always tell what's real.
The agent side got cheap the same way, and some of that's mine. Getting a content pipeline down to almost nothing was the whole goal, and it worked. If you've wired up a few agents and watched the per-run cost fall off a cliff, you know the feeling. Mine took a weekend.
The tooling that got my cost per run down is on npm, and it works the same for someone cloning a voice.
The FTC's 2025 numbers show what came out the other side. Americans reported losing 15.9 billion dollars to fraud, up 27 percent on 2024 and roughly 430 percent on 2020. Imposter scams took 3.5 billion of it across about a million reports, and business impersonators alone took about a billion.
That category has topped the FTC's list five years running. More people report being defrauded by someone pretending to be someone else than by anything else on it.
The patch is coming apart in court
Penske Media is suing Google over AI Overviews, and the case puts click-through down as much as 58 percent. The language in it is exact about what changed - Google went from a search engine that sends traffic to websites to an answer engine that removes the reason to click.
Then in late May a court in Munich ruled Google couldn't claim host-provider protection under the DSA for inaccurate AI Overviews. A court assigned liability for something a machine synthesized. Every vendor in the patch layer sits on that same defense, that they only pass things along and don't originate them.
I do AEO audits for publishers, so I watch this up close. Most of the work is making a publisher's material legible to systems that answer without linking. Every conversation ends up being about attribution, and it ends up there because nobody in the room can prove any of it.
One name, both sides of it
Deepfaked MrBeast ads have been running scams across YouTube for a while, using one of the most recognizable faces on the platform to take money from people who thought they recognized him.
In January, BitMine Immersion - Tom Lee's company, and as of that announcement the largest corporate holder of ether - put 200 million dollars into Beast Industries, partly to build a financial services platform on decentralized finance, sitting on an audience north of 450 million.
The same face is worth counterfeiting to a scammer and worth 200 million dollars to Tom Lee, inside the same twelve months.
Somebody's already writing the standard
While the suits proceed, ERC-8004 shipped. It's called Trustless Agents and it defines three onchain registries - Identity, Reputation, Validation. An agent gets a portable identity as an ERC-721 pointing at a registration file. The stated goal is letting agents find each other and trade trust signals across organizational boundaries with nobody central vouching for anyone.
It's live on Ethereum, Base, Polygon, Monad, and BNB Chain. The authors work at MetaMask, the Ethereum Foundation, Google, and Coinbase.
If you're wiring agents together across two companies right now, you're hand-rolling this, and in a year you probably won't be.
Google is a defendant in the publisher suits and a co-author of the agent trust standard at the same time. They're paying lawyers on one layer and shipping registries on the other.
What I got wrong about my own swing at it
Last year I went down a rabbit hole with GPT trying to learn this space, and what came back was a provisional patent draft for blockchain-based attribution in creator commerce. It wouldn't have survived examination, and working out why taught me more than filing it would have. The background section named the problem fine. Affiliate systems run on centralized links, cookies, and manual tracking, and those are prone to fraud, data loss, and disputes.
I got it in front of Gunderson Dettmer. Their read was that the claims were the easy part, and that a provisional with no product behind it is a filing fee rather than a position.
Good advice, because the claims wouldn't have survived anyway. Mine came down to six verbs. Assign an identifier, record engagement, pass it to checkout, match it, split the money, show a dashboard. That's affiliate marketing, and Amazon shipped it in 1996. Putting it on a chain doesn't make it an invention, and the best prior art against me was probably already sitting in the portfolio Circle went on to buy.
The technical problem was right there and I didn't claim it. Attribution state dies when it crosses a trust boundary. That's the actual hole, and if you're building anything that hands off between two systems that don't share a database, it's your hole too. It doesn't matter whether the thing crossing is a referral, a payment, or a claim about who somebody is.
Why it's worth something now
IBM spent ten years selling this to banks and insurers. An immutable record, a contract that pays out on its own, a way to trust a stranger with nobody in between. It mostly didn't stick, because banks already had middlemen they trusted and regulators standing behind them. They didn't need it badly enough.
Machines transacting with machines need it badly, and so does anyone whose face just became a monetizable asset. Which is what Circle bought when it went from owning almost no patents to holding the largest blockchain portfolio in the US, nearly a thousand of them from IBM. The framing is USDC protection. Most of what's in there was never about currency - supply chain verification, enterprise infrastructure, identity, insurance. We spent twenty years not building that, and Circle picked it up for the price of a patent portfolio.
I don't know what two more years looks like. I do know the patents that sat unused at IBM for a decade got bought the same year a face became worth 200 million dollars and worth counterfeiting, and that nobody in adtech built the thing in between.