Last week I laid out the Inversion Thesis: agents are becoming the interface, and websites are collapsing into API layers that only agents read. This week I want to follow that thesis into a place that should worry economists more than it currently does — the growing gap between real economic activity and the activity we actually measure.

Call it Ghost GDP: value that gets created, consumed, and never shows up in a national accounts ledger because the transaction that used to sit in the middle of it disappeared.

The disappearing middleman transaction

Here’s a small, concrete version of the problem. A person used to research a purchase by visiting five to ten websites, each of which served an ad, logged a session, and contributed a data point to somebody’s analytics and somebody’s ad revenue. Multiply that by billions of research sessions a year and you get a real, measurable slice of the digital economy.

Now the same research happens inside a single agent conversation. The agent reads structured sources, synthesizes an answer, and the person acts on it. The value delivered to the person is arguably higher. But the transaction layer that used to sit underneath that value never fires. The economic activity happened. The economic record of it didn’t.

This isn’t a rounding error. It’s the leading edge of a structural shift in how value gets created without generating the transactional exhaust that GDP accounting was built to capture.

Two kinds of businesses in this world

If Ghost GDP is real, it splits businesses into two categories, and the split has nothing to do with size or industry:

Toll-road businesses sit at a chokepoint agents have to pass through regardless of how the surrounding economy reorganizes — infrastructure, verified data, compute, identity and trust layers, the rails underneath the activity. They get paid whether or not the activity above them shows up in a measured transaction.

Storefront businesses depend on being the destination of a human-initiated visit. As more research and first-draft work moves inside agent conversations, storefront businesses lose the traffic that used to fund them, even while the underlying demand for what they offer stays constant or grows.

The uncomfortable part of this thesis isn’t that some businesses will struggle. It’s that the struggle won’t look like a demand problem. Revenue will decline while genuine interest in the category holds steady, because the interest is being satisfied one layer upstream, inside a conversation nobody can see or monetize the old way.

Where the tolls actually are

A few candidates that hold up:

• Verified, structured data that agents need to cite. If your content is the source an agent retrieves from rather than a page a human might visit, you’re upstream of the traffic collapse, not caught in it.
• Trust and citation infrastructure. Something has to tell an agent that a claim is current and real. Whoever builds that layer gets paid on every query that passes through it.
• Physical and logistical chokepoints that can’t be disintermediated by better retrieval. An agent can summarize trucking regulations instantly; it cannot drive the truck.

The honest caveat

I don’t think this is a clean, fast collapse — measured GDP has absorbed technology-driven mismeasurement before (the productivity paradox of the 1980s computing boom is the closest historical analog, and it took the better part of a decade to resolve in the data). What I do think is that the mismeasurement window is a real strategic hazard for any business whose model depends on being the destination of a human search, because the demand not showing up in your funnel isn’t demand that vanished — it’s demand that got served by something upstream of you.

Next week I want to take a step back from economics into something that looks unrelated but isn’t: what happens to creativity itself when a growing share of the content shaping AI training data is itself AI output. It’s the same homogeneity problem, one layer down.