An Agent Sent You a Customer. Your Analytics Called It “Direct.”

Someone signs up for your product today. Your analytics says the source was direct. No campaign, no referrer, no search keyword. They just… appeared.

Increasingly, that "direct" signup isn't direct at all. An AI agent recommended you — in a coding session, a chat, a research task — and the human acted on it. The agent did the selling. Your analytics gave the credit to nobody.

This is the attribution problem of the agent era, and it isn't a tracking bug you can patch. It's structural.

The buyer and the payer are two different actors

Classic attribution assumes one continuous human: a person clicks an ad, browses, and converts, carrying a cookie and a referrer the whole way. Every analytics tool ever built is shaped around that assumption.

Agent-driven discovery splits it in two. The agent reads about you — through your docs, an API, an MCP server, whatever machine-readable surface it can reach — server-side, with no cookie and no identity. Later, a human opens a browser and signs up. The actor who chose you and the actor who paid you are different, split across two events that share nothing your analytics can join on.

That gap is why the signup reads as direct. There is genuinely nothing in a standard session to tie the agent's read to the human's conversion.

Attribution is a spectrum, not a yes or no

The mistake is treating "can you attribute it" as a single question. There are really three cases, and they are wildly different in how measurable they are.

• The agent hands over a link, and the human clicks it. The clean case. The click carries a trail you can follow. Fully measurable.
• The agent and the human share a context — an agent working inside the user's own browser, say. The two events are close enough to connect. Measurable.
• The agent just says "use X," and the human shows up cold. No link, no shared session — they search your name two days later and sign up. There is no deterministic signal here. This is the hard one, and it may be the most common one.

The first two you can instrument honestly. The third you mostly can't — the best you get is asking "how did you hear about us" at signup, or a probabilistic correlation between an agent asking about a tool and a matching customer appearing soon after. Useful, directional, not proof.

Anyone selling you perfect AI attribution is lying

A wave of tools is about to promise full attribution of AI-driven revenue. Be skeptical. The cold-recommendation case is inherently lossy — you cannot deterministically tie a sentence an agent said in a conversation you never saw to a signup that happened in a different session days later. Pretending otherwise is how you end up with a confident dashboard that is quietly wrong.

The honest version is more useful: measure the deterministic slice precisely, label the probabilistic slice as probabilistic, and admit the invisible slice exists. A real number you can trust beats a big number you can't.

Why this is the bottom of the Agent Funnel

Pickrate measures the top of the agent funnel: when an agent has to choose a tool, does it choose you? That is your Pick Rate. But a great Pick Rate only matters if it turns into customers — and right now most teams can't see whether it does, because the conversions land in the direct bucket.

If agents are becoming the actor that picks tools — and we think they are — then agent-driven revenue is a channel. A channel you can't measure is a channel you'll under-invest in. You'll keep optimizing the surfaces you can see and ignore the one quietly sending your best-fit customers: the ones an agent already vetted.

We're instrumenting our own funnel this way first — honestly, deterministic slice only, no inflated numbers. When an agent picks us, we want to know whether it paid off. That is the whole point of measuring the funnel instead of just the score.

Curious whether agents pick you in the first place? Check your Pick Rate.