How personalized recommendations differ from other strategies

Recommender systems select products on different principles. The difference is whose data is used:

Strategy Basis Who it suits
Popular products Aggregate across the whole store Everyone equally, cold start
Similar items Attributes of the current product Product page context
Frequently bought together Transaction history Upsell and cross-sell
Personal The profile of one specific shopper Returning visitors

Personalized recommendations require accumulated data about the shopper — at minimum a few sessions
with views and actions.

How the personal profile is built

The algorithm assembles an affinity profile from events:

  • Product views — a weak signal, but high in volume
  • Add to cart — a strong signal of intent
  • Purchases — the strongest signal, weighted highest
  • Search queries — intent stated explicitly
  • Dwell time — scrolling down a product page beats a glance at it

From these events the system builds affinity vectors across brands, categories, price bands and
product attributes (colour, size, material).

Example affinity profile:
Brands: Nike 0.82, Adidas 0.61, Puma 0.38
Categories: Sneakers 0.91, Apparel 0.45
Price: $80–170 — the shopper's working range

Where to place personalized recommendations

Homepage — a For You widget above the fold or between promo blocks. High reach, and a good place
to personalize for returning visitors.

Category page (PLP) — You might like between listing rows or above them. Holds the shopper when
the current page is not quite what they wanted.

Product page (PDP) — Recently viewed and Picked for you below the description. Reduces the loss
of shoppers the current product did not convince.

Cart — Others buy this with it, based on the profile and the cart contents.

Email — a personal selection injected into a campaign sent from your own ESP, built from the
profile without any session context.

Tip: do not put personalized recommendations in a slot where another strategy performs better.
For a Similar items block on the PDP, content-based filtering is more precise. For cross-sell, use
frequently bought together. Keep personal selections for For You and Recently viewed.

Cold start: what to do with new users

There is no personal profile on a first visit. The algorithm accumulates data step by step:

  1. Visit 1 — contextual data only (device, traffic source, geolocation, time of day). Popular or
    trending products are shown.
  2. 3–5 product views — the first affinity signal. Recommendations start to adapt.
  3. 1 purchase — the profile becomes substantially richer. Personalization becomes accurate.

How to measure performance

For personalized recommendations, metrics belong in this order:

  1. Widget CTR — are recommended products being clicked
  2. Attributed revenue — revenue from orders that included a widget click, counted against an
    attribution window of 7–14 days
  3. Share of orders with a widget interaction
  4. RPV — revenue per visitor for shoppers who saw the widget, against a control group

Every one of these has to be compared in an A/B test: the personal algorithm against popular
products or against the previous algorithm.