Before diving into models and settings, let us understand what the service is and where it fits today.

The one-sentence definition

Recommendations AI is a fully managed Google Cloud service that uses machine learning to show each shopper products they are likely to want, based on your product catalog and on what shoppers do on your website or app. Google handles the data processing, model training, tuning and serving, so you do not need to build or operate your own recommendation system.

A note about the name

Google Cloud has reorganized its retail products over time. Recommendations AI began as its own service, later became part of the Retail API, and is now a capability inside Vertex AI Search for commerce. According to Google’s documentation, that product brings together recommendations, search and browse, and conversational commerce capabilities for retail websites and apps. Many guides, labs and training courses still use the older name “Recommendations AI”, and the console and API still use retail-focused naming, so you will see both. In this guide we use “Recommendations AI” for the recommendation capability, and “Vertex AI Search for commerce” for the wider product.

What it needs to work

Two kinds of information feed every recommendation model. The first is your product catalog: the list of products with names, categories, prices, availability and other attributes. The second is user events: records of what shoppers do, such as viewing a product, adding it to a cart or buying it. The models learn patterns from these two sources and then answer questions such as “what should we show this shopper on the product page?”

Recommendations are not decoration. In online retail they are often among the most valuable parts of the page.

The discovery problem

Large catalogs overwhelm people. A shopper who sees ten thousand products usually leaves without buying. Search helps those who know what they want, but many shoppers do not. They browse, get distracted and need guidance. Recommendations fill this gap by narrowing the catalog down to a handful of items that feel relevant.

What good recommendations do for a business

Why not build it yourself?

Some large companies build their own recommendation systems, but it is a serious undertaking. You need data pipelines, model research, training infrastructure, a fast serving layer, experiments and ongoing retraining as catalogs and tastes change. For most retailers, a managed service gets a strong system running much faster and with a smaller team.

A handful of terms appear in every recommendation project. Learn them once and the documentation will feel familiar.

flowchart LR
  C[Product Catalog] --> M[Trained Recommendation Model]
  E[User Events] --> M
  M --> S[Serving Configuration with Rules]
  S --> P[Prediction Request from Your App]
  P --> R[Recommended Products]
  R --> U[Shopper Clicks and Buys]
  U --> E
        

You do not need to understand the mathematics, but a simple picture of how recommendations are produced helps you feed the system well.

Learning from behavior

The basic idea is that shoppers who behave similarly tend to like similar things. If many people who viewed hiking boots also bought wool socks, the system learns a link between them. It does this at a huge scale, finding patterns among millions of events that no person could notice. It also looks at product attributes, so a brand-new item with little history can still be placed near similar items.

Sessions and context

Recommendations are not based only on long-term taste. The model also considers what the shopper is doing right now: which product they are viewing, what is in their cart and what they looked at earlier in the session. This is why the same shopper may see different suggestions on a product page than on the home page.

The life of a recommendation

sequenceDiagram
  participant S as Shopper
  participant A as Your Website or App
  participant R as Recommendations Service
  S->>A: Opens a product page
  A->>R: Write event: product viewed
  A->>R: Request recommendations for this shopper and product
  R-->>A: Ranked list of products and attribution token
  A-->>S: Shows suggestions
  S->>A: Clicks a suggestion
  A->>R: Write event with attribution token
        

The catalog is the foundation. Models can only recommend what they know about.

What goes into a catalog entry

Each product has a unique identifier, a title, one or more categories, and usually a description, price, availability, images and a link to its page. Optional attributes, such as brand, color, size, tags and custom fields, give models and business rules more to work with. The richer and cleaner the information, the better the recommendations and the more control you have.

Ways to load a catalog

Keeping the catalog fresh

Catalogs change constantly: new items arrive, prices change and stock runs out. Plan for a regular full import plus smaller, frequent updates for fast-changing fields such as price and availability. A stale catalog leads to recommending items that are gone or wrongly priced, which frustrates shoppers and wastes valuable screen space.

If the catalog says what you sell, events say what shoppers care about.

The common event types

How events get into the service

Visitor and user identifiers

Each event carries a visitor identifier, which tracks an anonymous browser or device, and optionally a user identifier for logged-in shoppers. Keeping these consistent lets the service link a shopper’s actions across a session and, when appropriate, across visits. Be careful that identifiers are stable but do not expose personal information.

Why historical data helps

Models perform better with enough events over enough time. Google publishes data requirements that differ by model type, covering the number and variety of events and the amount of history. If you are new to the service, loading past events can shorten the time to a good model. Check the current documentation for the exact thresholds for the model you plan to use.

Different spots in a shopping journey call for different kinds of suggestions. The service offers several model types, each designed for a specific question.

The exact list and the availability of each type can change, so check the current documentation. The core idea stays the same: pick the model whose question matches the moment in the shopper’s journey.

Optimization objectives

A trained model is the engine. Serving configurations are the steering wheel and pedals.

What a serving configuration does

A serving configuration connects a trained model to a named setup that your application calls. In it, you choose the model, set how many results to return, and add rules that shape the output. One model can be used by several serving configurations, for example one for the home page and another for emails, each with different rules.

Common business controls

Placements in practice

A clear order of work prevents wasted effort. Here is a practical path.

Recommendations AI is one part of a retail data and application landscape. Seeing the whole picture helps with planning.

flowchart TD
  subgraph Sources
    MC[Merchant Center]
    BQ[BigQuery]
    GCS[Cloud Storage]
    WEB[Website and Mobile Events]
  end
  MC --> CAT[Catalog in the Service]
  BQ --> CAT
  GCS --> CAT
  WEB --> EV[User Events in the Service]
  CAT --> TR[Model Training]
  EV --> TR
  TR --> SC[Serving Configurations]
  SC --> API[Prediction API]
  API --> FE[Storefront and App Placements]
  FE --> WEB
  EV --> DASH[Dashboards and Evaluation]
        

Where your own systems connect

A recommendation system is never “done”. You launch it, measure it and keep improving it.

Metrics that matter

A/B testing

The most trustworthy way to measure value is an experiment: show recommendations to one group of shoppers and the existing experience to another, then compare outcomes over enough time. This protects you from crediting recommendations for sales that would have happened anyway. Tools such as web analytics and experimentation platforms can help you run these tests.

flowchart LR
  A[Split Traffic] --> B[Group A: Existing Experience]
  A --> C[Group B: With Recommendations]
  B --> D[Compare CTR, Conversion and Revenue]
  C --> D
  D --> E{Clear Improvement?}
  E -->|Yes| F[Roll Out More Widely]
  E -->|No| G[Adjust Model, Rules or Placement]
  G --> A
        

Ways to improve

Real stores face awkward cases. Knowing them in advance saves frustration.

Shopper behavior is sensitive data. Handling it responsibly is both a legal duty and a trust matter.

Protecting data and access

Personal data and consent

Events can relate to identifiable people, especially when user identifiers are included. Follow privacy laws and your own policies: obtain consent where required, avoid sending unnecessary personal details such as names or email addresses in events, and use pseudonymous identifiers. Provide ways for shoppers to opt out of personalization, and have a process for deleting data when asked.

Even managed systems need watching. Early warnings prevent weeks of quietly poor results.

What to monitor regularly

Understanding the cost drivers helps you budget and avoid surprises.

Ways to manage cost

Several options exist for adding recommendations. The right one depends on your size, skills and goals.

flowchart TD
  Q1{Do you sell products online?} -->|Yes| Q2{Need results quickly with a small team?}
  Q1 -->|No| A1[Consider media recommendations or custom models]
  Q2 -->|Yes| A2[Recommendations AI in Vertex AI Search for commerce]
  Q2 -->|No| Q3{Do you have data science capacity and unusual needs?}
  Q3 -->|Yes| A3[Custom models on Vertex AI]
  Q3 -->|No| A2
        

Patterns capture what experienced teams repeatedly find works; anti-patterns capture what goes wrong.

Helpful patterns

Anti-patterns to avoid

An honest view of strengths, limits and where teams actually apply the service.

Real-world examples

A short checklist to carry into every recommendations project.

Recommendations AI is the recommendation capability, which is now part of Vertex AI Search for commerce. That product also includes search and browse capabilities and conversational commerce features. Older materials may still use the earlier name.

A product catalog and user events such as product views, add-to-cart actions and purchases. Models need enough data and history to train well, and the documented requirements differ by model type.

No. Google trains and serves the models. You need skills in data integration, web or mobile development and analytics, plus the ability to run experiments.

Training can take several days at first, depending on data. Gathering enough events and running a meaningful experiment takes additional time, so plan in weeks rather than hours.

Yes. Visitor identifiers track anonymous sessions, and recommendations use the current session’s behavior. Logged-in identifiers can add continuity where appropriate and permitted.

Keep availability up to date in the catalog, and apply filters in the serving configuration to exclude unavailable items.

It is a value returned with recommendations that you send back with subsequent events. It lets the service connect clicks and purchases to the recommendation that led to them, improving measurement and learning.

Pick the placement with the most traffic and clearest business value. Many retailers begin with “others you may like” on product pages or “recommended for you” on the home page, then add “frequently bought together” for the cart.

Run a controlled experiment and compare click-through, conversion and revenue between shoppers who see recommendations and a control group, over a period long enough to cover normal patterns.

The service provides encryption, access control and audit logging, but you remain responsible for consent, data minimization and respecting privacy laws. Use pseudonymous identifiers and avoid sending unneeded personal details.

Models need enough products and events to learn. Small sites may not reach the thresholds yet. Collect events, load history if available and consider simpler approaches until data volume grows.

Mainly by model training and prediction request volume, plus data and related services. Estimate your traffic, avoid redundant calls with caching, and check the current official pricing page.

Recommendations AI gives online retailers a managed way to show each shopper products they are likely to want. Fed by a product catalog and user events, Google trains retail-focused models that you connect to placements through serving configurations, shaped by business rules such as filters, diversity and price controls. It now lives inside Vertex AI Search for commerce, alongside search and conversational capabilities. Success depends less on clever settings than on solid foundations: consistent identifiers, complete event tracking, a fresh catalog, objectives matched to each placement, careful experiments, thoughtful privacy and steady monitoring. With those in place, recommendations can raise engagement, basket size and loyalty while remaining affordable and manageable.