Guides / Learning
Spectare personalises from the first visit and measures itself from the first visit. This guide explains what gets counted, how winners are called, and what you will see in your admin while it happens. No statistics background needed, though the statistics are the point.
There is no training period. The first visitor to arrive gets a personalised page, assembled from your content using the signals their visit carries. Learning runs alongside from day one, but it gates nothing: data improves the choices, it never delays them.
Every arrangement Spectare serves is recorded as a variant: which components, built from which atoms, shown to which audience at which stage. Three things are counted against it, per day: impressions (it was shown), continuations (the visitor went on to another page rather than leaving), and conversions (a goal you configured was reached). A conversion credits every atom the visitor saw on the way, not just the final page.
Evidence is windowed. When you edit an atom, the arrangements built from it start their evidence again from the edit: the old numbers described content that no longer exists, so they are not allowed to vouch for the new version.
If you turn the holdout on, a small share of visitors sees your default page instead of a personalised one. They are the control group. Lift is the measured difference between the two, on your real traffic, with the same goals. When Spectare quotes a number for what personalisation earned, this is where it comes from; without a holdout there is a comparison to nothing, which is how most tools quote lift.
The statistics are built so that watching the numbers every day cannot mislead you. Most testing tools assume you decide a sample size, wait, and look once; in practice everyone peeks, and peeking makes weak results look strong. Spectare uses confidence intervals that stay valid however often anyone looks. The practical consequence is honest and sometimes uncomfortable: a variant is called a winner only when the evidence genuinely clears the bar, and with modest traffic that takes time. You will see “not enough evidence yet” in your admin more often than a dashboard would like to admit. That is the feature.
A share of visits deliberately tries something different: a new arrangement, a component shape this audience has not seen, a different image. Exploration is how the system discovers anything worth exploiting, so it never fully stops, though it tapers as evidence accumulates. Exploration builds are Spectare’s own learning, so they are not metered against your plan.
When the numbers support a change to your settings, Spectare does not make it. It shows you a suggestion in Analytics with the evidence attached: what was measured, over how many impressions, and what it implies. You accept or dismiss. In learning mode (the default) your component rules are visible but read-only, because the point of the mode is that changes flow from evidence rather than hunches; switch to manual mode any time to take the wheel.
One thing matters more than everything else: configure conversion goals, and not only the final one. A goal that fires at signup tells the system almost nothing about the journey there; a goal on a brochure download or a pricing view gives it something to learn from every day. See the conversion tracking guide. Beyond that: review suggestions when they appear, check the Confidence page occasionally to watch evidence accumulate, and edit content freely, knowing the evidence resets honestly behind you.