Help Articles – Page 4
Adobe Commerce 2.4.6 left standard support on August 11, 2026. Adobe gives 2.4.6 customers a further year of extended support at no extra cost, covering quality and security patches for the core application, until August 31, 2027. A security-only period follows until May 31, 2028.
For planning, 2.4.5 left extended support on August 11, 2026, and 2.4.7 leaves standard support on May 31, 2027. Adobe publishes the dates in its Commerce lifecycle policy.
The platform does not switch off, and the store keeps taking orders. What ends is the vendor's support for that version: no new security fixes, no compatibility updates for the PHP, Java or .NET runtime underneath, and no help from the vendor's support desk.
From then on, every new vulnerability is yours to patch, mitigate or accept. For merchants, that also affects PCI DSS v4.0.1 requirement 6.3.3, which expects patches for critical vulnerabilities to be installed within one month of release. With no vendor patches, meeting it relies on workarounds your team builds and maintains.
- Precision@k: of the top k items shown, the share the customer actually wanted.
- Recall@k: of all the items the customer would have wanted, the share that made it into the top k.
- Hit rate@k: the share of customers for whom at least one of the top k items was a hit.
- NDCG@k: similar to precision, but gives more credit when the hits appear near the top of the list.
- Coverage and diversity: how much of the catalogue gets recommended at all, and how varied each list is.
These are useful for comparing models offline. Business outcomes such as conversion, basket size, repeat purchase, returns and retention, confirmed with live experiments, tell you whether recommendations are working for customers.
An eCommerce recommendation engine decides which products appear for each shopper: similar items and frequently bought together on product pages, add-ons in the cart, the order of search and category results, personalized homepages, and products in emails and notifications.
A good one does three things well. It reads complete, current product data, including price and stock. It respects business rules such as margin, promotions and items that should never appear together. And in B2B, it works within each account’s catalogue, pricing and reorder patterns. Increasingly it also feeds AI shopping assistants that build a full basket from a single request.
- For customers: less searching, more relevant choices, and discovery of products they would not have found.
- For the catalogue: exposure for new and long-tail products as well as bestsellers.
- For the business: personalization across thousands or millions of customers, larger baskets through relevant add-ons, and more reasons to come back.
The advantages depend on good product data and on measuring the right outcomes. A system tuned only for clicks can narrow what people see and work against repeat purchase.
Collaborative filtering struggles with new items, and content-based filtering tends to repeat what someone already likes. A hybrid system combines them, for example by blending their scores, switching between them depending on how much data is available, or using one to fetch candidates and another to rank them.
Most large production systems are hybrids. Target’s stack, for example, layers embeddings, affinity models, next-item prediction and similarity models before ranking and re-ranking the results.
Content-based filtering looks at the items themselves. If a shopper buys relaxed-fit jeans in a dark wash, it recommends other items with similar attributes. It works for a new product as soon as its attributes are filled in, which makes it useful when there is little purchase history.
Its weakness is that it tends to recommend more of the same, and it is only as good as the product data behind it. Incomplete or inconsistent attributes lead directly to weak recommendations. Most systems combine it with collaborative filtering.
- Amazon: “frequently bought together” and “customers who bought this also bought,” plus its Alexa for Shopping assistant.
- Netflix: personalized rows of titles on each member’s home screen.
- YouTube: the videos shown next to the one you are watching.
- Spotify: personalized playlists and playlist continuation.
- Google: the Discover feed, personalized Search results and AI-written notifications.
- Target: “Recommended for you,” “Complete the look,” “Your top deals” and an on-site shopping assistant.
Recommendation engines also make less visible decisions, such as the order of search results, which coupon a customer receives and which products appear in an email.
- Collaborative filtering: recommends what similar people liked or bought.
- Content-based filtering: recommends items that resemble what someone already likes, using product attributes and descriptions.
- Hybrid: combines the two, which is what most production systems do.
- Context-aware: adjusts for location, season, device or time of day.
- Knowledge-based: uses rules and constraints, such as compatibility or a customer’s contract catalogue.
- Deep learning: neural models, such as two-tower and sequential models, that learn from very large amounts of behavior.
- Generative: AI that produces the recommendation itself, such as a summary, a plan or a basket.
An online recommendation engine suggests items to each visitor on a website or app: the products in a “customers also bought” row, a personalized homepage, the order of search results, the next video, or the products in an email. It works from what the visitor has browsed and bought, what they have told the site directly, the context they are in, and what similar visitors chose.
Recommendation engine, recommendation system and recommender system all mean the same thing. In 2026 the definition is widening to include AI assistants that write the recommendation itself, such as a meal plan with a shopping list or a full basket for a project.
Clicks and add-to-cart rates are fast and useful, but a click doesn’t prove a recommendation was good. People click on whatever sits at the top of a list, and systems trained only on clicks learn to serve whatever is easiest to engage with.
A stronger approach combines three things: offline tests to screen ideas, controlled live experiments to confirm them, and longer-term measures such as repeat purchase, returns, retention and customer satisfaction. Several RecSys 2026 speakers, including Thorsten Joachims of Cornell and the team at Lyft, focused on optimizing for those longer-term outcomes.
A recommender learns from history, so a first-time visitor or a newly listed product gives it little to work with. Common fixes include using product attributes and descriptions, context such as location or season, popular items, and simply asking the customer.
Some systems live with cold start permanently. Google described its Discover feed as being in “eternal cold start” because it focuses on fresh content that has no history yet. Plain-language preference profiles and conversational assistants are also being used to close the gap faster, because a customer can say in one sentence what would otherwise take many clicks to infer.
Collaborative filtering looks for patterns across many customers. If people who bought one product also tend to buy another, a new customer who buys the first is shown the second. It needs no understanding of the products themselves, only behavior.
Its weak spots are new or rarely bought items with little history, and content generated for a single person, which has no crowd behind it at all. That is why most systems combine it with content-based methods that use product attributes, descriptions and images.
Most recommendation engines work in stages. First they gather signals: what the customer has browsed, searched, clicked and bought, what they have said they want, the context they are in, and what similar customers chose. Then they retrieve a few hundred likely candidates from a catalogue that can run to millions of items, rank those candidates for the person, and finally apply business rules such as stock, margin, variety and promotions.
The funnel is so well established that Amazon described using the same stages at RecSys 2026 for a different job: choosing which AI model should power each shopping feature. Newer systems add a generation step, where AI writes the recommendation itself, such as a summary, a plan or a basket.
Margaret Mitchell’s opening keynote at RecSys 2026 argued that people who approve AI actions all day start approving quickly, and that the AI then learns to produce whatever is easy to approve. Her suggestions included:
- Planting a known error from time to time, sometimes called a canary, to check that reviewers still catch it.
- Tracking outcomes over time, such as errors found after the fact, in addition to approval rates.
- Having people return to doing the work themselves periodically, so the skill to spot a problem does not fade.
It also helps when AI changes arrive as staged drafts, with a clear record of what changed and who or what made the change.