Recommendation Algorithms — What Each One Picks

Every recommendation widget is a frame; what goes in it is decided per visitor, at the moment it renders, by the
algorithm you pick. This is what each one draws on, and what it needs before it can return anything.

Three videos go deeper into what the algorithms pick and how to shape it:

Video: Product recommendations: what to show each visitor (2:02) — it plays at the top of Recommendations — Where to Start.

Pick the page first — it changes the algorithm, not just the placement

The picker asks for a page before it offers you anything: Home, Category,
Product, Cart, Thank you, Search,
404.

That is not a filter on where the widget appears — the Target step decides that. Behavioural algorithms
anchor to something different on each page: the product being viewed on a Product page, everything
recently viewed on Home, the cart contents on Cart. The same algorithm on two pages is two different questions.

The groups

Two things on the picker itself are worth knowing. Each card carries a scope chip naming what it anchors to on the page you picked — based on the item being viewed, from recently-viewed category — so you can see the anchor without opening anything. And a card that cannot work yet says so on its right-hand side: “Needs a product feed”. The Has data filter hides every algorithm in that state at once.

Personalized & Behavioral

• Personalized Recommendations — the visitor’s profile plus the behaviour of similar visitors.
• Most Popular — the most viewed items in the period you choose.
• Best Sellers — highest purchase count.
• Most-Added to Wishlist — most often saved to favourites.
• Trending Now — see below; this one is rate-of-change, not a ranking.
• Inspired by Past Purchases — what people who bought the same things went on to view and buy.

Trending Now — and whose trend

Trending Now compares today’s activity against the item’s own weekly average, so a steady
best-seller does not trend and a small item catching fire does.

• Trending by picks the signal: Views, Purchases, Add to cart, or Wishlist.
• From picks whose trend it is: the whole site, the visitor’s categories, or
  the visitor’s interests. The last two are personal trending — what is taking off in the corner of
  the catalogue this person actually cares about.

Site-wide is the default on purpose: a scoped trend depends on the visitor, so it is no longer safe to reuse in an
email or an ad. Give that up deliberately, not by accident.

Cross-Sells, Co-Views & Upsells

• Cross-Sells — the bought-together relations, chosen under its own Based on menu.
• Others Who Viewed Also Viewed, Others Who Viewed Also Added to Cart,
  What Others Viewed Then Bought — the viewed-together family.
• Up-Sells — from an up-sell catalogue you upload.

The co-view cards carry a From scope of their own: anywhere, the same category, or
the same interests — so “also viewed” can be held inside the aisle the visitor is standing
in. One relation, one card that names it: the bought-together flavours live on Cross-Sells, the viewed-together
flavours on the co-view cards.

Visitor History

• Recently Viewed — what they looked at, most recent first.
• Buy it Again — their own purchases.
• Your Wishlist — what they saved.

Catalog Events

• Price Dropped — items whose price fell.
• Back in Stock — items that returned to stock after selling out.
• New in Stock — items that arrived.

Back in Stock

The engine records the exact moment an item’s stock crosses in either direction, automatically, from your
normal feed imports. There is nothing to configure on the feed.

The window can be since the visitor’s last visit, or a fixed span from yesterday up to
three months — the same longer windows now available on Price Dropped.

It is offered on all seven page types, and Cart and Thank-you are deliberate: a restock of something a visitor
abandoned is one of the strongest recovery triggers there is. It combines with every filter and fallback, and with
Skip items the visitor already… — though leave has purchased before switched off if you
are writing a repurchase reminder.

Similar Items (AI), and Similar Content (AI)

Every other algorithm on this page needs traffic first: an item nobody has viewed, carted or bought is invisible to
them. Similar Items (AI) reads the catalogue itself — titles, brands, descriptions, categories
(for articles: title, topic, author, description) — and computes each item’s nearest neighbours by meaning — within its own category, which is both what a widget should show and what keeps a large catalogue affordable.
A brand-new product gets recommendations on day one.

Switch it on: Catalog settings → data settings → Similar items (AI). One
switch covers both catalogues, products and content. Use it: the Similar Items (AI) card in
the recommendation editor, under Cross-Sells; Similar Content (AI) for articles.

• It runs nightly, and only over items that were added or changed — so a stable catalogue
  costs very little after the first pass.
• It uses AI credits and respects the account’s daily AI cap; if it hits the cap it resumes
  the next night.
• Until you switch it on the card serves nothing and the widget’s fallback covers — the widget is never
  broken, just not yet AI-fed.
• Available on Home, Product, Cart and Thank-you for products; Home and Article for content.

Matches Their Last Search

Every other algorithm on this page recommends from behaviour — what this visitor interacted with, what co-occurred with it, what is popular. None of them can answer a phrase, and a phrase is the most explicit statement of intent a visitor ever makes.

Matches their last search reads the phrase Personyze keeps on the visitor’s profile — typed into your own site search, or carried in from a search engine — and matches it against your catalogue’s own words: title, brand and short description for products, title, topic and description for articles.

• It answers a phrase the first time anyone types it. There is no nightly job and no phrase list to accumulate, which matters because search tails are long and the phrase that counts is usually new.
• Matching is natural language, not a query syntax, so a visitor typing C++ or an unbalanced quote cannot steer or break it. A phrase that matches nothing returns nothing.
• Relevance decides who is in the list; personalization only reorders it. Someone who typed a phrase is owed items matching that phrase, not the items you would have shown anyway. Inside the matching set, the usual freshness nudge and per-visitor price band apply.
• It needs a phrase. Until the visitor has searched, the widget renders nothing — so the recommendations editor binds the page group Searched on this site for you, and says so on the widget.

The full picture — what is captured, how it differs from the Keywords rule, and how to target on it — is in targeting and recommending on the last search.

Content recommendations

The content catalogue has the same shape, and three algorithms of its own worth naming:

• What Readers Went On to Convert On — a read→goal collaborative filter: what readers of
  the same articles engaged with before reaching a goal.
• Content You Liked — this reader’s own favourited and saved articles; the per-person
  twin of Most Liked Content.
• Inspired by Reached Goals — seeded by the goals this visitor reached; the content twin of
  Inspired by Past Purchases.

Every algorithm, by the name the picker shows

The groups above explain how the engine thinks. This is the index: every algorithm you can actually choose, under the name it carries in the picker. Which ones are offered depends on the page you picked first — an algorithm anchored to the current product is not offered on a home page — and several appear in more than one place with the wording adjusted to the context.

Popular right now

No personal history needed — these work on the first page view, which is what makes them the usual fallback.

Personalized to the visitor

The behavioural model: what this visitor is likely to want, given what people like them did.

What other visitors did with the same item

Co-occurrence. Each needs other visitors to have done the pair of things first, which is why a quiet catalogue falls back.

Up-sell

Deliberately higher-priced than the anchor, for margin rather than fit.

The visitor’s own history

No model at all — a list of what this person did. They need that person to have done it, so they render nothing for a first-time visitor.

Content

The article-catalogue equivalents. “Extra” is a comment here, not an add-to-cart.

Not a recommendation list

Output — how many, and what happens when it comes back short

• Recommendations to show — how many slots.
• Fallback algorithms — a second (and third) algorithm behind the first. A union is usually a
  top-up rather than an only-when-empty fallback, so several can serve at once. The
  QA bar will tell you which one actually filled
  each slot.
• Fill empty cells — off by default. Off means only what the algorithm returned is shown, and
  an empty result leaves an empty block. The fallback above is separate and applies either way.

Fine-tune the algorithm

Gentle multipliers on the chosen ranking — Boost by profit margin, Boost new
items, Reserve a slot for new items. The algorithm still leads; these nudge.

These apply to this widget only. The footer shows the account-wide setting live —
“Account-wide blend now: …” — so you are never tuning one widget blind to the
account it sits in. See the score
blend.

Related

• Category & interest recommenders — the same engines, recommending a category or topic instead of an item.
• Score blend & presets — the account-wide weights behind every “Recommended” ordering.
• Recommendation fallbacks
• Filters on recommendations
• Interests — what the interest-scoped options draw on.
• Recommendations aren’t appearing