Three generations of method
The first generation was rules: a marketer writes audience definitions by hand — returning visitor, paid traffic, Northeast — and attaches an experience to each. Rules are legible but brittle; they cover only the audiences someone thought to define, and they lean on identity signals most traffic never provides. The second generation was model-picks: a machine-learning system chooses among marketer-built experiences using profile and history data. Better coverage, but the profile prerequisite reappears — the model has little to say about the visitor it can't recognize.
The current generation is in-session behavioral scoring: the visit itself is the signal, scored live, so the anonymous first-time visitor — the majority of any site's traffic — is personalized as fully as the known customer. On top of it sits a newer axis: agentic operation, where the segments, variants, and placements are built and tuned by AI agents through a tool surface rather than clicked together by hand.
What actually gets personalized
In practice, a small set of page elements carries most of the value. The hero — headline, image, primary claim — matched to the visitor's apparent intent. The offer: discount framing for the price-sensitive, differentiation for the comparison shopper. Product recommendations ranked from live behavior rather than global bestsellers. Calls to action whose urgency matches how deep in the visit the visitor is. And social proof — reviews, logos, guarantees — surfaced for the hesitant. Two visitors on the same URL, seconds apart, can each get the page that fits where they are.
How to evaluate website personalization software
Four questions sort the market faster than any feature grid. What share of traffic can it act on — does it need a resolved identity, or does it work on anonymous visitors? Does the decision land before the page paints, or do visitors see the flicker? Is lift measured against an always-on holdout, or asserted by a dashboard grading its own homework? And can your AI agents operate it through a real tool surface, or only your team through the vendor's UI? The comparison pages apply these tests vendor by vendor.