Somewhere in the last few years, the internet stopped being a place people visit and became a place machines read. Crawlers index it. Retrieval systems summarise it. Agents transact across it. Humans still show up — but as a share of total requests, the human internet is shrinking toward a rounding error.

Advertising has not absorbed what that means, because advertising is built almost entirely on the assumption that the opposite is true.

Every layer assumes a person

A bid request implies a person to bid on. A viewability standard implies eyes. Frequency capping implies a someone to cap. Brand safety implies a reader forming an impression. Attribution implies a decision that can be influenced. Every one of those primitives encodes the same assumption: that a request is a person, and non-human traffic is the exception you filter out.

Invalid traffic detection was designed for that world. It asks: is this request fraudulent? That was the right question when automation on the open web was mostly adversarial. It is the wrong question now, because most automated traffic isn't fraud. It is infrastructure — legitimate, declared, and doing useful work on behalf of a human who is somewhere else entirely.

We have a category for machines pretending to be people. We do not have a category for machines acting openly on a person's behalf. That gap is where the next decade of supply-chain argument will happen.

What breaks first

Not spend — measurement. Denominators get unreliable before budgets do. When an unknown and growing share of requests never had a person attached, the metrics built on top drift in ways that look like performance changes rather than definitional ones. Publishers see traffic that doesn't monetise. Buyers see efficiency that doesn't convert. Both sides reach for optimisation levers when the problem is arithmetic.

The second thing to break is intent. If a person asks an agent to research and buy, the agent generates the requests and the person makes the decision. The impression and the intent have come apart. Advertising has never had to price that separation before.

The useful response

Not panic, and not another filter. Three things:

• Measure the composition, not just the volume. If you cannot say what share of your supply is human-attached, you cannot price it. Most buyers today cannot.

• Stop treating declared automation as fraud. Adversarial and agentic traffic need different handling. Collapsing them into one number destroys the information in both.

• Price the path, not the impression. The more requests decouple from people, the more value sits in knowing how supply was assembled rather than how many times it rendered.

The industry spent fifteen years getting good at buying attention at scale. The scarce thing now is not attention. It is knowing which requests ever had a person behind them.

[Draft note for review: this argument needs one hard number to anchor it — the human share of total requests, with a source we're willing to stand behind publicly. Everything else here is structural and defensible without a citation.]