How to spot a high-intent visitor from behaviour

What are the five signals?

They are worth listing in order of how much they actually tell you, which is not the order most dashboards put them in.

Return visits first. Somebody on their third visit this week has a reason. Nothing else on this list survives as well when you check it against who eventually bought.

Then the page. Pricing, plans, checkout, book a demo, contact: pages nobody opens idly. A person reading one of those is further along than a person who has read nine blog posts.

Then depth within the visit. Four pages in one session is engaged. Nine is either serious or lost, and which one it is usually shows in whether the pages are getting closer to the pricing page or further from it.

Then dwell time, which is weaker than it looks and is covered below.

Then the source. Somebody who clicked a paid ad for your product category asked a commercial question a few seconds ago. That is a real signal, and it decays fast.

What does a scoring model look like in practice?

Concretely, here is the one IntellQ ships with. Ten rules, 0 to 100, and every workspace can edit the weights and the thresholds. It is written out rather than described because a scoring model nobody can read is a scoring model nobody trusts.

Why are the weights arbitrary?

Because they are, and any vendor who tells you theirs are derived from science is describing a model trained on somebody else's customers.

The weights are a starting guess. What matters far more is the band boundary: the line above which a person gets a human's attention. A model that ranks correctly is useful even if every number in it is wrong, because ranking is all you asked it to do. Two visitors, which one first, is the actual question.

So tune the bands before you tune the points. If your operators are greeting people who turn out to be researching a school project, the hot line is too low. If the board is never hot, it is too high.

What does not predict intent?

Four things that feel like signals and are not, each of which will be on a dashboard somewhere.

• Time on site by itself. A tab left open over lunch outscores a person who read your pricing page carefully for ninety seconds. Always pair dwell with what they were reading.
• Pageviews by itself. Somebody lost in your documentation generates more pageviews than somebody who went straight to checkout.
• Country and device. They change how you should answer and they say nothing about whether this person is buying.
• Scroll depth on a short page. On a page that fits the screen, everybody is at 100 percent and the signal is noise. It is a real signal on a long page and only there.

What should happen when the score goes up?

Something specific, or do not compute it. A score that only sorts a list is a decoration.

The two useful responses are different in kind. One is to move that person up the queue so a human greets them first, which costs nothing and is almost always right. The other is to open a conversation automatically with a proactive trigger, which is an interruption and is wrong more often than people expect.

Set the trigger on the conjunction, never on the score alone. On the pricing page, more than forty seconds, second visit, is a rule with a reason behind it. A bare score over 70 fires on whoever happens to be reading a lot today.

IntellQ evaluates trigger rules in the browser rather than on the server, against state the page already has, so a rule can combine path, dwell, referrer, campaign, visit number, scroll depth and exit intent without a network round trip every few seconds per visitor.

How do you know the model is right?

One measurement, run monthly, and it is not the one most teams run.

Take the visitors your model called hot last month and work out what share of them converted. Do the same for warm and for cold. If hot does not convert at a visibly higher rate than warm, the model is sorting noise and the weights need work.

The number to watch after that is what share of conversions came from visitors the model never called hot. That is the cost of the line being too high, and it is the expensive error, because those people were ready and nobody spoke to them.