AI Product Recommendations for Magento 2

H2: Your Store Recommends Products Like It's 2010 — And Shoppers Notice

Walk into any conversation about ecommerce conversion and someone will quote the same statistic: Amazon attributes roughly 35% of its revenue to its recommendation engine. McKinsey has long pegged personalized recommendations as a driver of 10–30% of ecommerce revenue. Salesforce research has found that shoppers who click a recommendation are dramatically more likely to convert and spend more per order.

Now look at what most Magento 2 Open Source stores actually ship: the native "Related Products," "Up-sells," and "Cross-sells" blocks. These are not algorithms. They are manual lists that a merchandiser has to wire up product by product, by hand, in the admin. On a catalog of a few hundred SKUs that is tedious. On a catalog of tens of thousands it is impossible, so the blocks sit empty or stale. The result is a product page that recommends nothing relevant, a cart that misses the obvious add-on, and a homepage that treats a returning customer exactly like a first-time stranger.

The painful part is that the technology to fix this is no longer exotic. Adobe Commerce ships Adobe Sensei–powered Product Recommendations — but Sensei is Commerce-only, and Commerce licensing starts around $22,000 per year and climbs fast with GMV. For the roughly two-thirds of Magento stores running Open Source, that door is closed. They are left choosing between an enterprise license they cannot justify and a manual related-products workflow that does not scale.

The consequence of doing nothing is quiet but expensive: every product page that shows irrelevant or empty recommendations is a missed cross-sell, a lower average order value, and a shopper who bounces to a competitor whose site "just gets" what they want. This article is about closing that gap on Open Source — using the same kind of AI that powers the big players, at a fraction of the cost.

It is worth being honest about why the native blocks fall short, because the limitation is structural rather than a matter of effort. Magento's related products, up-sells, and cross-sells are stored as explicit relationships you create in the admin. They have no concept of "this product is like that product." If you add a thousand new SKUs next quarter, every one of them ships with empty recommendation blocks until a human gets around to linking them — and most never do. Worse, the moment your catalog changes, your hand-built links drift out of date: discontinued items linger, seasonal products recommend out-of-season companions, and nobody has the bandwidth to audit it all. The blocks that were supposed to drive discovery quietly become dead weight that you ship to every visitor.

H2: What Is AI Product Recommendations (and Why Your Magento 2 Store Needs It)

MageRa AI Product Recommendations is a recommendation engine for Magento 2 that replaces hand-built related-products lists with semantic, AI-driven suggestions. Instead of you telling Magento "show product B on the page for product A," the engine understands your catalog and figures out which products are genuinely similar or complementary.

It does this with embeddings. Each product — its name, descriptions, key attributes, and categories — is converted into a numeric vector using OpenAI's text-embedding-3-small model. A vector is just a long list of numbers (1,536 of them) that captures the meaning of the product. Two products that are conceptually similar end up close together in this vector space, even if they share no keywords. A "lightweight trail running shoe" and a "breathable off-road sneaker" describe nearly the same thing; embeddings recognize that, where a keyword match would not.

When a shopper lands on a product page, the engine finds the nearest vectors to that product and surfaces them as recommendations. That is fundamentally different from Magento's native capability, which has no notion of similarity at all — it only shows what a human explicitly linked.

The psychology behind why this works is simple: relevance reduces friction. When the next product a shopper sees is obviously related to what they are already looking at, browsing feels effortless and discovery feels personal. That is the same mechanism behind Netflix's "Because you watched…" and Spotify's recommendations — reduce the effort of finding the next good thing, and people consume more.

The typical impact is measurable rather than magical: published benchmarks for recommendation widgets cluster around a 2–6% lift in click-through rate and a 3–8% lift in average order value when relevant recommendations replace generic or empty blocks. On a store doing meaningful volume, single-digit AOV percentage points add up quickly.

Consider two concrete scenarios. First, a fashion retailer with 8,000 SKUs across dozens of categories: a shopper viewing a navy linen blazer should see other lightweight summer tailoring, not a random winter parka that happens to share the word "jacket." Embeddings capture that nuance automatically across the entire catalog, something no merchandiser could maintain by hand. Second, a home-goods store running a flash sale: the merchant wants a specific promoted lamp to appear in the "More Like This" block on a handful of high-traffic pages. With manual pin overrides, that takes thirty seconds, and the AI fills the remaining slots with genuinely similar products — the best of both worlds.

**When is this not the right fit?** If you have a tiny catalog of a dozen highly curated products where a human merchandiser already hand-picks every cross-sell with care, an AI engine adds little. It also is not a substitute for good product data — if your descriptions are empty, the embeddings have little to work with. The engine shines when you have a real catalog (hundreds to hundreds of thousands of SKUs), decent product text, and not enough hours in the day to merchandise it all manually.

H2: Key Features of MageRa AI Product Recommendations for Magento 2

• More Like This (semantic PDP recommendations) — Finds the genuinely most-similar products to the one a shopper is viewing, using vector similarity rather than manual links. The block fills itself, even on a catalog you have never merchandised.
• Customers Also Viewed (collaborative filtering) — Learns from real behavior, scoring products that shoppers view, add, and buy alongside the current item. It backfills with semantic similarity until enough behavioral data accumulates, so the block is never empty.
• Recommended for You (personalized picks) — Builds a "taste vector" from a visitor's recent activity and surfaces personalized products in the cart. First-time visitors see popular products instead — a graceful cold-start, not a blank space.
• Category-centroid cold-start — A brand-new product with no embedding yet borrows the average vector of its category, so it can be recommended (and receive recommendations) from day one instead of waiting for the next batch.
• Behavioral re-ranking — Blends pure semantic similarity with click → cart → purchase signals, nudging proven performers up the list.
• A/B testing and analytics — Split traffic between the AI engine and native related products, then compare CTR and purchases on a built-in dashboard to prove the uplift before you commit.
• Manual pin overrides — When merchandising or a promotion demands it, force a specific product into any widget; pins always appear first.

The "so what?" behind every one of these features is the same: less manual merchandising, more relevant product discovery, and a measurable lift in the metrics you actually report on.

H2: How to Set Up AI Product Recommendations in Magento 2

Step 1: Prerequisites

• Confirm you are on Magento 2.4.6+ and PHP 8.3 / 8.4 / 8.5.
• Create an OpenAI account and generate an API key with access to the embeddings endpoint. Budget for embedding cost — text-embedding-3-small is inexpensive, typically a few cents per thousand products.
• Make sure Magento cron is running (bin/magento cron:install). Embedding generation is a background job, not a page-load operation.

Step 2: Installation

Install via Composer and run the standard upgrade sequence:

composer require magera/module-ai-recommendations
php bin/magento module:enable MageRa_AiRecommendations
php bin/magento setup:upgrade
php bin/magento setup:di:compile
php bin/magento setup:static-content:deploy -f
php bin/magento cache:flush

This creates the embedding, tracking-event, and override tables automatically.

Step 3: Core Configuration

• Go to Stores > Configuration > MageRa Extensions > AI Recommendations.
• Set Enable AI Recommendations to Yes.
• Open the OpenAI Embeddings group and paste your API Key — it is stored encrypted at rest, never in plain text.
• Leave Embedding Model as text-embedding-3-small and Vector Dimensions as 1536 unless you have a specific reason to change them.
• Under Storefront Widgets, switch on the three widgets you want per store view.

Step 4: Advanced / Fine-Tuning

• In the Recommendation Engine group, set Vector Search Backend to Auto. It will use OpenSearch k-NN if your store has it and silently fall back to in-database PHP cosine similarity if not — so it works on any Open Source install.
• Keep Category-Centroid Cold-Start and Behavioral Re-ranking on.
• Tune Minimum Similarity Score (start at 0.15) and Rate Limit (requests/minute) to balance result quality against OpenAI cost.

Step 5: Generate Embeddings and Test

• Run the first full embedding pass (ideally off-peak):

bin/magento magera:airecommendations:generate-embeddings --queue-all

• Watch Catalog > AI Recommendations > Analytics Dashboard until the embedded count covers your catalog.
• Open a product page and confirm the More Like This and Customers Also Viewed blocks render with relevant products, then check the cart for Recommended for You before going live.

H2: MageRa AI Product Recommendations for Magento 2

• Built for Open Source performance: embeddings are generated via cron, queue, and CLI — never during a storefront page render — and results are cached, so the shopper-facing pages stay fast. Vectors are stored compactly in a MEDIUMBLOB column, and only products whose text actually changed are re-sent to OpenAI (content-hash detection), keeping ongoing cost minimal.
• Differentiator 1 — Adobe Sensei–class results without Commerce: the same embeddings-plus-vector-search approach that powers enterprise engines, delivered on Magento Open Source for the price of an OpenAI key plus the extension, instead of a five-figure annual Commerce license.
• Differentiator 2 — works with or without OpenSearch: where the OpenSearch k-NN plugin is available it is used for fast vector search; where it is not, the module computes cosine similarity over its own table. No mandatory infrastructure, no extra unsupported Composer packages.
• Differentiator 3 — privacy by design: product text is sanitized to strip emails and phone numbers before anything leaves your store, no customer PII is sent to OpenAI, visitor identifiers are stored only as hashes, and the API key is encrypted.
• Compatibility: Magento 2.4.6+ (Open Source and Commerce), PHP 8.3–8.5.
• https://magera.ca/magento-2-ai-recommendations.html →

H2: Tips and Best Practices for AI Recommendations

• Tip 1 — Fix your product data first. Embeddings are only as good as the text they read. Products with rich, accurate names and descriptions produce sharp recommendations; empty descriptions produce vague ones. A quick content cleanup before your first embedding run pays off immediately.
• Tip 2 — Run the first pass off-peak, then relax. The initial full-catalog embedding is the only heavy job. After that, saving a product re-queues just that product, and the nightly cron handles it. You should rarely need a manual run again — reach for rebuild-index only after a major catalog or attribute overhaul.
• Tip 3 — Tune the minimum similarity score deliberately. Set it too high and widgets go empty on niche products; set it too low and you surface weak matches. Start at 0.15, watch the dashboard CTR, and adjust in small steps.
• Tip 4 — Give "Customers Also Viewed" time, and trust the backfill. Collaborative filtering needs behavioral data to shine. Until it accumulates, the widget backfills with semantic similarity, so it is never empty — but its real strength appears after a few days of traffic.
• Tip 5 — Prove the lift with A/B testing before going all-in. Enable A/B testing at, say, a 50% AI traffic share for a couple of weeks, then compare CTR and purchases between the control and AI variants on the dashboard. Walking into a stakeholder meeting with your own uplift numbers is far more persuasive than quoting an industry average.
• Tip 6 — Mind your OpenAI budget knobs. The rate limit and batch size settings exist precisely so you can pace API usage. For a very large catalog, a lower rate limit spreads the initial pass across several nightly cron runs instead of one expensive burst. Because re-embedding is incremental, steady-state cost after that first pass is usually negligible — you are only paying for the handful of products that changed each day.
• Tip 7 — Keep the widgets honest with placement and copy. A recommendation block buried below the fold rarely earns clicks. Place "More Like This" near the gallery or add-to-cart area, give each widget a clear heading, and let the dashboard CTR tell you whether shoppers are actually engaging. Small placement changes often move the needle more than algorithm tuning.

H2: Conclusion

Most Magento Open Source stores still recommend products with hand-built lists that do not scale and cannot understand similarity — leaving cross-sells, average order value, and returning-customer relevance on the table. MageRa AI Product Recommendations closes that gap with OpenAI embeddings and vector search, delivering More Like This, Customers Also Viewed, and personalized picks, complete with a cold-start for new products, behavioral re-ranking, A/B testing, and manual overrides. Expect the kind of measurable lift — low-single-digit CTR and AOV gains — that compounds across every product page and cart.

CTA: Stop merchandising by hand and let your catalog recommend itself. MageRa AI Product Recommendations for Magento 2 brings enterprise-grade personalization to Open Source — install it and run your first embedding pass today.

https://magera.ca/magento-2-ai-recommendations.html →

FAQ

Q1: Do I need OpenSearch or Elasticsearch for this to work? A1: No. The engine uses OpenSearch k-NN vector search when it is available, but it automatically falls back to computing cosine similarity over its own embeddings table in PHP. That means it works on any Magento 2.4.6+ Open Source store without extra search infrastructure.

Q2: Will calling OpenAI slow down my product pages? A2: No. Embeddings are generated entirely in the background via cron and CLI — never during a page render. The storefront only reads pre-computed vectors and cached results, so shopper-facing performance is unaffected. OpenAI is only contacted when products are (re-)embedded.

Q3: How much does it cost to run, and is my data safe? A3: Beyond the one-time extension price, your only ongoing cost is OpenAI usage, which is low — text-embedding-3-small costs a few cents per thousand products, and re-embedding is incremental (only changed products). On privacy, product text is sanitized to remove emails and phone numbers before it is sent, no customer PII ever leaves your store, visitor IDs are hashed, and your API key is stored encrypted.

Internal Links

• /blog/magento-2-ai-visual-search (AI Visual Search — pair semantic recommendations with image-based product discovery)
• /blog/magento-2-voice-search (Voice & Conversational Search — extend AI-driven discovery to voice and natural-language queries)
• /blog/magento-2-agentic-commerce-ucp (Agentic Commerce (UCP) — expose your catalog and recommendations to AI shopping agents)