Checked on August 27, 2026.
An e-commerce recommendation system selects products based on the user’s context and the specific placement. A practical launch starts not with a neural network, but with one module—such as “Customers Also Bought”—and a measurable baseline. A model is useful only within a framework that accounts for availability, price, catalog constraints, and the A/B test result.
In brief: Choose one placement and one target action, set up logs for impressions and responses, and compare the model with popular-item or product-based rules. In production, you usually need candidate generation, scoring, and re-ranking. Offline metrics help select versions, but only a properly run online experiment can confirm business impact.
Contents
- Where to show recommendations
- What data you need
- How the system works
- How to get started without a complex model
- Cold start and anonymous visitors
- How to evaluate quality
- How to run an A/B test
- What to monitor after launch
- Pilot plan
- FAQ
- How AI Dawn implements recommendations
- Bottom line
Where to show recommendations
One store solves several different problems. You should not mix them in a single metric.
| Placement | Context | Typical goal |
|---|---|---|
| homepage | user and session | open a relevant category or product |
| product detail page | current SKU | find a substitute or add-on |
| cart | basket contents | offer a compatible add-on |
| email/push | history and trigger | bring the user back to purchase without unnecessary contacts |
First, define the placement, available space, latency, business constraints, and the success event. A click is convenient as a quick signal, but an order, margin, and return rate better reflect value.
What data you need
A minimal log includes the user or session ID, time, placement, the list of actually shown SKUs and their positions, clicks, cart activity, order, and cancellation. You also need separate catalog snapshots: category, attributes, price, availability, and whether the item is allowed to be sold.
A view is implicit feedback, and the absence of a click does not prove dislike: the item may have been below the fold or not shown at all. TensorFlow Recommenders explicitly treats views as an implicit signal and separates retrieval and ranking (official guide). Therefore, save exposure data, otherwise the model learns from an unknown display policy.
How the system works
Google describes a common three-stage architecture: candidate generation, scoring, and re-ranking (Google ML).
- Retrieval quickly selects hundreds of candidates from the catalog.
- Scoring estimates the probability of the target action for a context-product pair.
- Re-ranking applies availability, compatibility, diversity, freshness, brand limits, and other rules.
It is better to filter prohibited and unavailable items before retrieval and again before display. Prices and stock levels change faster than the model, so they need a fresh source.
How to get started without a complex model
A strong baseline is popular products by category, frequently bought together, manual bundles, and recently viewed items. It is cheap, explainable, and shows whether you have enough traffic for personalization.
The next level is content-based similarity using product attributes. Collaborative filtering uses co-behavior, while a two-tower model is convenient for large catalogs. Complexity is justified only if the new version consistently beats the baseline and stays within latency limits.
Cold start and anonymous visitors
For a new user, use session context, traffic source, category, and region instead of inventing a profile. For a new SKU, use the product page, brand, price, and text or visual features. Mix personalized candidates with popular and fresh ones; keep the exploration share limited and measurable.
Do not group people by questionable identifiers. Retention periods, consent, and access to behavioral data must comply with internal policy and applicable law.
How to evaluate quality
For offline evaluation, use a time-based split: train the model on past data and test it on a later period. Useful metrics include Recall@K, Precision@K, NDCG@K, coverage, diversity, and the share of unavailable items. Calculate metrics by placement, new and returning users, categories, and devices.
Historical data was created by an older algorithm, so it contains positional and selection bias. Research shows that offline recommendation evaluation can produce paradoxical conclusions because of the closed feedback loop (RecSys paper). It serves as a quality filter, not a promise of sales growth.
How to run an A/B test
Randomize at a level that prevents spillover: usually the user, sometimes the session. Before launch, define the primary metric, guardrails, minimum detectable effect, duration, and stopping rules. The control gets the current baseline, and the test gets the new policy.
In addition to clicks, check add-to-cart, paid orders, margin, returns, latency, and complaints. Do not end the test at the first nice-looking chart, and do not change the set of variants midstream without a new experiment version.
What to monitor after launch
- catalog availability and freshness;
- the share of requests with recommendations and fallback;
- latency and errors at each stage;
- position distribution, coverage, and repeatability;
- CTR, orders, margin, and returns by segment;
- feature drift and gaps between offline and online signals.
Store the model version, candidates, rules, and the displayed list. This makes it possible to reproduce the decision and roll back safely.
Pilot plan
- Choose one placement and one metric owner.
- Lock in the baseline and acceptance criteria.
- Check the logs for impressions, responses, and the catalog.
- Launch rules and a simple personalized version in shadow mode.
- Check the filters, latency, and fallback.
- Run an A/B test with guardrails.
- Document the decision to scale or stop.
FAQ
How much data do you need for a recommendation system?
There is no universal threshold. You need enough data for the chosen placement, segment, and measurable experiment; for rare events, it is better to start with rules and a content-based approach.
Can you make recommendations without registration?
Yes. Use items from the current session, page context, and category popularity, while following consent and data retention requirements.
What is better: an off-the-shelf service or your own model?
An off-the-shelf service speeds up the pilot, while your own system gives you more control over data and logic. Compare integration, total cost of ownership, latency, data export, and experimentation capabilities.
Does personalization guarantee sales growth?
No. The effect depends on assortment, placement, baseline, traffic, and experiment quality. It is validated on data from a specific store.
How AI Dawn implements recommendations
AI Dawn starts with one recommendation block: it records the current algorithm, business metric, catalog, constraints, and acceptance criteria. Then the team can:
- collect impression, response, and order events;
- implement baseline, retrieval, ranking, and re-ranking;
- integrate the service with the storefront and experimentation framework;
- set up monitoring, fallback, and handoff procedures.
Conclusion
A recommendation system for an e-commerce store is a controlled display policy, not just a model. Start with one placement, a reliable exposure log, and a strong baseline. Offline metrics will help you filter out weak versions, and the launch decision should be made based on a preplanned A/B test and guardrails.