Personalized Product Recommendation

What it is & the business problem it solves

Personalized Product Recommendation gives every customer their own ranked list of products, such as "Recommended for you" in a newsletter or on the homepage. It is part of Inspire.

It solves two problems:

  • Relevance: instead of one bestseller list for everyone, each customer sees products that match their own viewing and buying behaviour.

  • Business control: you decide what may be shown. You can leave out products a customer already bought, avoid repeating the same products, limit how many items from one category appear, push margin or new products, and pin campaign products to fixed positions.

The personal lists come from the recommender engine (see Recommender Engine). This module applies your business rules, tops up lists with a fallback and delivers the result.

The module comes in two versions in the cockpit, one per recommender model:

  • Personalized product recommendation transformer reads its lists from the transformer-powered engine. Since June 2026 the recommender is powered by a transformer model, and new customers start with this version by default.

  • Personalized product recommendation reads its lists from the classic engine (collaborative filtering).

Both versions are the same module in two flavours: the same business rules, the same delivery and the same way of configuring. They differ only in the engine underneath and in a few settings, listed in the section "Which recommender model does this module use?".

How it works

  1. Take the engine's list. For each customer, Inspire reads the raw personal list from the recommender engine (100 products by default from the classic engine; 10 by default from the transformer engine).

  2. Apply exclusions. Products are removed according to your rules: inclusion and exclusion criteria, previously bought products, previously recommended products.

  3. Shape the list. Diversity limits ("max 2 per category"), boosts and fixed positions are applied.

  4. Rotate (optional). Between runs, Inspire can rotate through the list so that customers do not see the same top products every time.

  5. Top up with fallback. If too few personal products remain, the list is filled from a fallback: a generic list such as the most popular products.

  6. Deliver. The final list is written to the customer profile in CDP (CM's customer data platform) or CXP (CM's customer experience platform). From there you use it in segments, workflows, MMC email campaigns and web personalization. The Recommendations API serves each customer's latest list plus a generic fallback list for customers without one.

Runs are incremental by default: only customers whose input changed are recalculated, not everyone.

Which recommender model does this module use?

The recommender engine behind this module is either the classic collaborative-filtering engine ("customers who engaged with X also engaged with Y") or the transformer engine, a neural network that reads the order of a customer's shopping journey. Both are described on the page Recommender Engine. The model is not a setting inside this module. You choose it by which version of the module you use:


Personalized product recommendation transformer

Personalized product recommendation (classic)

Engine it reads

Transformer Recommender

Base User Item Recommender

Default for new customers

Yes

No

Raw list length by default

10 products (set on the engine)

100 products (set on the engine)

Number of recommendations (final list)

1–100, default 3

1–500, default 3

Fallback Method

popularity, popularBuy, popularView

popularity, popularBuy, popularView, itemAttribute

Fallback Configuration: Item Attribute, Sort Order

Not available

Available

Fallback for Users Without Recommendations

Not available in the cockpit

Available (Expert)

Statistics

None predefined; use the event viewer and the fallback tags

Fallback Rate, Fallback Distribution, Fully Fallback Distribution and product-attribute distributions

Events in the cockpit event viewer

Personalized Product Recommendation Transformer

Personalized Product Recommendation

Extra prerequisite

The Data module's transformer loader must be on and scheduled

None

Packaging

Enterprise

Enterprise

All other settings in the tables below are the same in both versions.

Example use cases

Use case

Scenario

Channel

Required products

Weekly "picked for you" newsletter block

A fashion retailer adds 4 personal products (illustrative) and excludes items bought in the last 90 days (illustrative)

Email

CDP or CXP, Webtracking, Product Feed, MMC Email Campaigns

Homepage "For you" carousel

A DIY retailer shows 12 personal products (illustrative), with at most 3 per category

Website

CDP or CXP, Webtracking, Product Feed, Web Personalization (Recommendations API)

Campaign push

A pet-food webshop boosts its own brand and pins a seasonal product to position 1 during a promotion

Email, website

As above

Re-engagement with fresh products

A retailer excludes products already recommended in the last 2 weeks (illustrative), so repeat mailings stay varied

Email

CDP or CXP, Product Feed, MMC Email Campaigns

A/B test of the two recommender models

A home-improvement retailer shows a personal homepage carousel and A/B-tests lists from the transformer version of this module against lists from the classic version

Website

CDP or CXP, Webtracking, Product Feed, Web Personalization (Recommendations API), A/B Testing module (optional)

Two audiences from one engine

A retailer runs two instances of this module on one engine: one for newsletters (exclude bought items) and one for the website (boost new arrivals)

Email + website

As above

What you can configure

You can configure the settings below yourself in the cockpit (Inspire's front-end tooling for adjusting the business rules of your predictions). Settings marked Expert need expert rights. CM staff and customer admins can change expert settings; to get expert rights, ask your CM contact person.

Setting (cockpit name)

What it does

Default

When to change

Who can change

Recommender Schedule

When the module runs (a cron schedule in Amsterdam time)

Not scheduled: the module stays idle until set

Always set at go-live, after the recommender engine it reads from has run

Standard

Number of recommendations

Final list length per customer (1–500 in the classic version, 1–100 in the transformer version)

3

Match the number of slots in your email or web block

Standard

Rotation of Recommendations

Rotates through the list between runs (1–5; 1 = no rotation)

1

For frequent mailings to the same audience

Expert

Full Run Required

Recalculates every customer on each run instead of only changed customers

Off (incremental)

After major rule changes. Note: enabling Exclude Previously Recommended Items always forces full runs.

Expert

Email Filter (Enable Email Filtering)

Only creates lists for customers with a valid email address

Off

For email-only use cases

Standard

Fallback Method

How short lists are topped up: popularity, popularBuy (most bought), popularView (most viewed) or, in the classic version only, itemAttribute (ranked by a product feed column)

popularity

popularBuy for conversion-focused emails; itemAttribute to rank by, for example, stock level

Standard

Fallback Configuration (Number of Fallback Recommendations, Lookback Period (Days), Item Attribute, Sort Order)

Size of the fallback list (1–10000); days to look back for popularBuy / popularView (1–365); column and ascending/descending order for itemAttribute (Item Attribute and Sort Order exist in the classic version only)

100 items; 30 days

Larger lists help when business rules are strict

Standard

Fallback for Users Without Recommendations (classic version only)

Also delivers a list made entirely of fallback products to customers who have no personal list

Off: these customers are dropped and get no list

Switch on if every customer in a segment must receive a list

Expert

Item Inclusion Criteria / Item Exclusion Criteria

Product rules built in a criteria builder, for example only "in stock" or no gift cards

Off

Almost always: exclude out-of-stock and non-sellable items

Standard

Exclude Previously Bought Items

Removes products the customer bought: always, within a period (1–365 days), or except a set number of allowed repeat purchases (0–15). Can work on a similar product level (for example same model) and on selected products only.

Off

For durable goods; leave consumables (food, cosmetics) allowed

Standard

Exclude Previously Recommended Items

Removes products that were among the top 1–15 of earlier recommendations within a period, optionally unless the customer engaged with them, optionally only those recommended by selected modules (Module Selection)

Off

For frequent mailings to the same audience

Standard

Limit Similar Recommendations

At most 1–15 products per group (for example per category)

Off

To keep lists varied

Standard

Limit Products by Specific Features

At most 1–15 products matching a condition (for example "on sale")

Off

To control the product mix

Standard

Boost by Condition / Boost by Numerical Feature

Raises products matching a condition, or with a high value in a numerical column such as margin; strength 0–10

Off

To favour own brands, new arrivals or margin

Standard

Fixed Position Recommendations

Places a product matching a condition at a fixed position (1–15)

Off

Campaign or sponsored products

Standard

Personal Recommendation Filtering (Enable Minimum Personal Recommendations, Minimum Number of Personal Recommendations)

Requires a minimum number of genuinely personal products (0–100) in each list

Off

When lists made mostly of fallback are not wanted

Standard

Delivery Settings (Delivery Method, Batch Method, Event Type ID, CXP Definition ID, item and user information)

Where results go: None or Batch. Batch Method is CDP or CXP; file export to storage or a database exists for special set-ups. Each module only needs the Event Type ID (CDP) or the CXP Definition ID (CXP). The CDP account and access token are configured once at profile level by CM.

Batch to CDP; customer email and Webtracker client ID included

Set by CM at setup

Expert

Advanced/expert settings (User Inclusion Criteria, Expert Recommendation Filtering, Expert Boost, Expert Fixed Position Recommendations, System Resources Configuration)

Free-form expression rules and computing resources

Off; 1 CPU, 8 GB

Only with CM guidance

Expert

Prerequisites & required CM products

Requirement

Why

CDP (CM's customer data platform) or CXP (CM's customer experience platform)

The platform that holds your customer data and receives the lists. The Batch Method in the delivery settings decides whether lists go to CDP or CXP.

Standard event types on CDP: the Retail & Ecommerce adapter

Sets up the standard event types (Purchase, Add to cart and so on). It is included in onboarding: there is nothing extra to buy or configure. What matters is that events use these standard event types; custom event types are ignored.

Standard event types on CXP: Templates

The Retail & Ecommerce adapter does not apply to CXP. On CXP you use Templates to set up the equivalent events.

Webtracking (CM's website tracker)

Product views

Purchase lines (order lines from the webshop or back office)

Purchases weigh most and drive "exclude bought" rules

Product Feed

Product names, images, categories and attributes used in rules. Product IDs in events must match the feed's item ID.

Recommender engine

Required (see Recommender Engine): the Transformer Recommender for the transformer version of this module, the Base User Item Recommender for the classic version. The engine must be scheduled before this module.

Data module: transformer loader (transformer version only)

Enable Transformer Loader must be switched on and a Transformer Loader Cron set (CM setup). The engine's prediction starts after this loader finishes. The loader also requires the Product Feed to be enabled in the Data module.

Email channel

MMC Email Campaigns (lists via CDP or CXP profile properties)

Website channel

Web Personalization (via merge tags or the Recommendations API)

This module works with both CDP and CXP data.

Packaging tier

Available in the MMC Enterprise package only (not in Basic, Advanced or Pro). Inspire as a whole is available from the Pro package upward; this module is one of the Enterprise-level modules.

Data requirements

  • Events: these come through the recommender engine. The classic engine uses product detail views and purchase lines by default; the transformer engine uses product detail views, add-to-cart, remove-from-cart and purchase lines in time order. Exclusion rules for previously recommended items also use the module's own earlier results.

  • Transformer version, technical minimum requirements (minimums for the model to work, not a promise of quality): the engine trains only on customers who have at least one purchase, and whose journey has more than 3 events spread over at least 2 separate sessions. A product needs at least 1 engaged customer to be learned. For prediction, every customer with at least one event receives a raw list; customers without events get nothing from the engine, and this version does not create a list for them.

  • New and delisted products (transformer version): products that appear after the last training run are unknown to the model until the next retrain, and a product delisted after training can still be suggested until then. Retrain regularly and use an Item Exclusion Criteria rule such as "in stock = false".

  • Product feed attributes used in your rules (category, brand, stock, margin) must be filled.

  • Cold start: customers without any interactions get no personal list. They are dropped, unless Fallback for Users Without Recommendations is on; then they receive a fallback list. Popularity fallback also covers customers whose personal list becomes too short after business rules.

  • Coded rule: a run fails if more than half of the customers would end up with fewer products than the configured Number of recommendations.

  • Technical minimum requirements. There is no coded minimum number of customers, orders or history. The technical conditions that do apply: the recommender engine must have produced lists (customers with no recent activity get none), and the run stops with an error when more than 50% of the customers would end up with fewer products than Number of recommendations (a fixed limit that cannot be changed in the cockpit).

How to check it works

  • Preview: when you edit a draft variant, the preview runs on a sample of about 5% of customers, so you can check the effect of rules before publishing.

  • Event viewer: open the Personalized Product Recommendation events (classic version) or the Personalized Product Recommendation Transformer events (transformer version) in the cockpit. Each sample customer shows the recommended products with tags "Contains fallback recommendations" and "Only contains fallback recommendations".

  • Statistics: in the classic version the statistics are Fallback Rate, Fallback Distribution and Fully Fallback Distribution (how much of the output is fallback), plus distributions of the recommended products over product attributes such as category or brand. Statistics are set up per module instance. The transformer version has no predefined statistics; use the event viewer and the fallback tags.

  • Fallback share: if many lists show fallback tags, raise the engine's Number of Recommendations or relax the business rules.

  • A/B test: compare click-through and revenue of the two versions with the A/B Testing module.

  • Customer profile: open a test profile in CDP or CXP and check the recommendation property and its "Updated At" timestamp.

Limitations & common pitfalls

  • Idle after setup. Recommender Schedule (and the engine's schedule) start as "not scheduled".

  • Run fails with "too few recommendations". If more than half of the customers would get fewer products than Number of recommendations, the run fails. Relax rules, lower the number, or enlarge the fallback.

  • Missing customers. Customers without a personal list are dropped unless Fallback for Users Without Recommendations is on.

  • Stale lists on the customer profile. (Verified for CDP.) Lists are not cleared when a customer drops out of a later run. Use the "Updated At" property in your segment to ignore old lists, for example older than 14 days (illustrative).

  • Boosts multiply. The boost strength is a multiplier on the score, and several boost rules do not add up. Test with the preview.

  • Rules on the wrong product level. If your catalogue has variants, write criteria for the same level (variant or parent) that the recommendations use.

  • New lists are not instant. A run must age briefly and API responses are cached.

  • High fallback rate (transformer version). The transformer engine stores only 10 products per customer by default. After your exclusions there may be too few left, and lists fill up with popular products. Raise the engine's Number of Recommendations.

  • Profile flooding (transformer version). The engine's Inference User Scoping is off by default, so every customer with an event is predicted and delivered on every run. Turn it on for large audiences.

  • Nothing happens after setup (transformer version). Both the engine's Training Schedule and this module's Recommender Schedule start as "not scheduled", and the Data module's transformer loader must be on and scheduled.