What is a conversion?

The definition is a decision

A conversion is not a category of event; it is whichever event you decided to count. That choice determines what every optimisation system in your stack will pursue.

The common failure is choosing a shallow event because it has volume. Optimise to form fills and you will get form fills, including worthless ones. Optimise to a trial start and you will get trials that never convert. Machine learning systems are extremely effective at delivering exactly what they are scored on.

Choosing the right one

• Deep enough to indicate value, so that more of them is genuinely better.
• Frequent enough to optimise against, since a rare event provides no signal.
• Measurable reliably, ideally server-side — see conversion tracking.
• Stable in definition, because changing it invalidates historical comparison.
• Where the deep event is too rare, the workable compromise is to optimise on a shallow proxy while monitoring the deep one, and re-baseline when the relationship between them drifts.

Conversions without a click

The standard model assumes a click, a session and an event fired from a page. A growing share of outcomes has none of those.

An agentic conversion arrives as an order with no preceding web behaviour. An AI-influenced purchase arrives through search or direct, with the influencing exposure invisible. In both cases the conversion is real and the path is not observable.

The practical response is to measure conversions server-side against your own system of record, and to answer the causal question — what caused them — with a holdout rather than with a referrer.