What is session intelligence?

How is session intelligence different from analytics?

Analytics answers questions about populations after the fact: how many, from where, how often. Session intelligence answers one question about one person, now: what is this visitor doing, and what is in their way.

Both are useful and neither substitutes for the other. A funnel report telling you that 60% abandon at step three is how you learn there is a problem. It cannot tell you that *this* visitor is abandoning right now, which is the only knowledge you can still act on.

Session recordings sit between the two and are often confused with this. A recording is session-level, but it is retrospective and requires a person to watch it — which means it scales to the few dozen sessions someone has time for, out of the hundreds of thousands recorded.

What does session intelligence produce?

A compact, structured account rather than a pile of events: what the visitor appears to want, how far along they are, whether they are frustrated, and what they are likely to do next.

The compression is the engineering problem. A few minutes of browsing produces thousands of raw events, and nothing useful can be done by handing all of them to a model — it is expensive, it is slow, and most of it is noise. Compiling them into a short account first, in ordinary code, is what makes reasoning about every session affordable rather than about a sample.

It is also where privacy is decided. An account of behaviour describes what someone did; it need not contain what they typed, who they are, or where they connected from. A system that compresses first can be built so that the sensitive material never leaves the edge, because it was never part of what got sent.

What is session intelligence used for?

Deciding whether this visitor needs anything, and if so what. It is the input to intent prediction and, downstream of that, to any intervention — which is why its quality matters more than the sophistication of whatever reasons over it.

A good account of a session and a modest model outperform a poor account and an excellent one, because the model can only reason about what it was told. Most of the difficulty in this field is in the description, not in the inference.