Every session, scored on whether it meant it.

Every analytics tool answers the question “was this traffic any good?” the same way: it counts the conversions that came from it. That works beautifully at scale and falls apart everywhere else.

Consider a channel that sent 400 sessions and produced one sale. Is that a 0.25% conversion rate, or is it one lucky sale from traffic that was worthless? With one data point you cannot tell, and you will not be able to tell next month either — you will have two data points. Meanwhile the spend continues.

The problem is that conversions are rare events, and rare events are terrible units of measurement at small sample sizes. But the sessions themselves are not rare. Four hundred sessions is a lot of behaviour. The information is there; counting just throws almost all of it away.

Intent scoring uses the behaviour rather than the outcome. Every one of those 400 sessions gets evaluated against the pattern your converting sessions have historically shown, and gets a number.

Now the question changes shape. Instead of “one sale from 400 sessions”, you have a distribution: perhaps 12 sessions scoring above 70 and 340 scoring below 20. That is a channel bringing a small amount of real interest. Or perhaps nothing above 30 at all — which tells you, from a single month and one sale, that this channel is not bringing buyers and the sale was incidental.

The same reasoning runs the other way. A source with no conversions at all but a healthy tail of high-scoring sessions is usually a landing-page or offer problem, not a traffic problem — and that is a completely different fix from turning the channel off.

Only behaviour observed on your own site. There is no third-party intent data, no company identification and no cross-site profile.

• Reading depth. How far through the page the visitor actually got, per page, rather than whether the page loaded.
• Attention. Active time versus idle time, and whether the tab was switched away from. A page open for nine minutes in a background tab is not nine minutes of interest.
• Path shape. Which pages, in what order. The move from a feature page to pricing means something different from the reverse.
• Pace and depth of visit. Pages per session, session duration, and how the visit accelerated or stalled.
• History. Whether this browser has been here before, and what its earlier sessions looked like.

The model learns which combinations of these preceded conversion on your site. That last part matters more than the signal list: the same behaviour means different things on a checkout funnel and on a considered B2B purchase, which is why a general-purpose engagement score cannot do this job.

The scoring model is a well-understood statistical one rather than something exotic, and that is a deliberate constraint rather than a limitation we are apologising for.

It trains reliably on the amount of data a normal business has, instead of needing the volumes a deep model would demand. Its outputs can be explained signal by signal, so you can see why a score moved. And when it is wrong, it is wrong in ways you can inspect — which is the difference between a tool you come to trust and a black box you quietly stop opening.

It is also validated before it is ever shown. Each trained model is tested against data it has not seen, and if it cannot separate converting from non-converting sessions well enough, it does not go live. The reporting says the model is not ready. That is a worse day for us and a better one for you.

The category overclaims, so we would rather draw these lines ourselves than let them be assumed.

• Not a prediction about a person. It is a comparison across a population. A session scoring 90 may never return. The number is for ranking, not fortune telling.
• Not incrementality. A channel bringing high intent may be harvesting demand you created elsewhere. Brand search is the classic case. Scoring tells you where intent showed up, not who caused it.
• Not identity. No cross-site tracking, no company lookup, no third-party intent feed. It cannot tell you which company visited.
• Not causal. The model learns which behaviours accompanied conversion. It cannot tell you that manufacturing more of a behaviour would produce more conversions.
• Not zero-data. It needs some conversion history to learn from. Far less than the alternatives, but not none.

Google Analytics does have predictive metrics, and they are good at what they are for. But a property only becomes eligible once at least 1,000 returning users have triggered the event and 1,000 have not, inside 28 days — roughly 33,000 to 50,000 returning sessions a month at typical ecommerce conversion rates. It predicts purchase or churn specifically, at user level, for the next seven days, and it exists mainly to build remarketing audiences.

Privacy-first analytics — Plausible, Fathom, Umami, Matomo — report traffic accurately and cheaply and do not attempt scoring at all. If clean traffic reporting is what you need, they are the better buy and we will say so.

B2B intent platforms — 6sense, Demandbase, Bombora and others — score accounts, not sessions, using third-party intent data and company identification. Different data, different unit, enterprise pricing, and only useful if you sell to companies.