E-commerce
March 12, 2025
Do you want to increase average order value and conversions without multiplying paid campaigns? Smart product recommendations show the right item at the right point in the journey: product page, cart, sometimes search or conversation. This guide explains what a “smart” recommendation is, which signals it uses, how to deploy it on Shopify, and how to measure it without relying on magic percentages found on anonymous infographics.
Estimated reading time: 15 min
Summary
Definition and promise
A smart product recommendation adapts to the visitor's profile, the current session, or the cart: it is not just a fixed « best-sellers of the week » carousel. It can combine business rules (margin, stock, season) and models that learn from clicks and purchases. The goal is twofold: help people choose faster and increase cart value when the suggestion is truly complementary.
For more details on AI and personalization, our article AI product recommendation for more sales explores the strategic framework and market sources.
In practice, « intelligent » does not mean « an incomprehensible black box »: a good implementation often blends explicit rules (never pair one SKU with another) and scoring learned from data. Your added value is deciding what must remain under human control and what can vary automatically.
Why move from static to adaptive
The Shopify blog on AI in e-commerce describes how search, merchandising, and conversational assistants combine to bring customers closer to relevant products. On a personal level, shoppers are increasingly accustomed to experiences that remember their preferences: a store that always shows the same three items unrelated to the current browsing session can seem behind the times.
« Companies that excel at personalization generate 40 percent more revenue from those activities than average players. »
McKinsey, The value of getting personalization right or wrong is multiplying
Recommendations are only one lever among others (email, loyalty, support), but they are visible on every product page visit: that makes them a good candidate for measurable tests.
Signals: behavior, catalog, context
Systems generally use:
Behavior: views, add-to-carts, searches, collection journeys. On Shopify, the web pixels and standard customer events (for example
product_viewed,product_added_to_cart) feed these signals when they are collected correctly.Catalog: categories, attributes, variants, prices, availability. Without clean product data, AI does not infer that an accessory is compatible.
Context: device, traffic source, current campaign, current cart. An “add-on to cart” suggestion must take into account what has already been selected.
Purchase history remains very valuable for signed-in customers, provided you comply with your legal obligations and your data policy.
Block types and objectives
Block type (example) | Main objective | When to use it |
|---|---|---|
Similar / alternatives | Help compare or replace | Product page, search |
Add-ons / cross-sell | Increase cart value | Product page, cart |
Frequently bought together | Transactional social proof | Product page |
For you / personalized | Returning visitor or identified customer | Home page, customer account |
Trends / new arrivals | Discovery for cold traffic | Home page, emails |
Algorithms: content, collaborative, hybrid
Approach | Idea | Strengths / limitations |
|---|---|---|
Content-based filtering | Similar product by attributes (material, use, range) | Works well early with little history; sensitive to tag quality |
Collaborative | "Those who liked A also liked B" | Strong with volume; cold start for new SKUs |
Hybrid + rules | Model + exclusions (out of stock, margin) | The most common in e-commerce production |
The Shopify article on AI personal shoppers (see the e-commerce AI framework above) illustrates how, for the same customer question, one can combine search intent and preferences to recommend products: a logic similar to combining search + recommendations on the store.
Internally, document who "owns" the logic: marketing (priority to new products), finance (priority to margin), or product (compatibility and brand image). Without explicit arbitration, default tools often optimize for short-term clicks, not necessarily the assortment strategy.
Where should recommendations be placed on the store?
Product page
It's the most natural place: the visitor is in decision mode. Alternate between «similar items» and «complete your outfit / your kit» depending on your world.
Cart and checkout
Cross-sell should stay light: an inexpensive accessory, a warranty, a consumable. Avoid adding a dozen lines that push the order button out of the viewport on mobile.
Homepage and collections
The «for you» blocks need enough signals or a fallback (trending items) for new visitors. Otherwise, you always show the same best sellers.
Email and follow-ups
Post-purchase journeys and abandoned cart campaigns benefit from suggestions related to the viewed item or the cart. Consistency with the site's prices and stock is essential.
Catalog quality: a prerequisite often overlooked
Before adding a sophisticated engine, check: unique titles, variants, attributes (compatibility, size, material), representative images, up-to-date stock. An “intelligent” recommendation that points to an incomplete product page or an out-of-stock item undermines trust faster than a static carousel.
Visual consistency: sharp, consistent thumbnails in the carousel.
Language and tone: use the same terms as on the product page for materials and uses.
Variant management: if the recommendation displays a color, make sure it is available.
If you structure your data with metaobjects and metafields, you make useful matches between complementary products easier.
Implementation on Shopify
Choose the solution : native features, Shopify app, or chatbot with recommendations (see below). Evaluate cost, maintenance, support, and data compliance.
Define placements : start with the product page and cart, measure, then expand.
Exclusion rules : out of stock, sensitive product lines, products already in the cart, items you do not want to promote together.
Tests : A/B test on the number of items, the block title, and mobile order.
Data : make sure the pixels and the necessary events are properly fired to feed the models.
Mobile first : validate scrolling, thumbnail size, and the block position relative to the buy button.
International : if you sell in multiple currencies or languages, check that the suggestions respect local catalogs.
Post-launch review : schedule a review at D+7 and D+30 to adjust exclusions and titles.
Metrics and interpretation of results
Indicator | Question | Action if anomaly |
|---|---|---|
Recommendation CTR | Do people click on the suggestions? | Review the title, placement, relevance |
Click conversion | Do the clicks lead to purchase? | Pricing, stock, or product page issue |
Average order value | Are orders with recommendations higher? | Check whether the suggested products are too cheap |
Attributed revenue share | What share of sales goes through the blocks? | Compare periods before / after with the same traffic |
Cross-check with Shopify Analytics for actual sales and margins, not just clicks.
To avoid congratulating yourself on a high CTR for products that bring in nothing, segment by margin category or by collection: a recommendation that pushes “loss leader” products can inflate volume without feeding profitability. Also compare visitors from paid search and organic search: the behavior is not the same.
Bias, diversity, and customer experience
A model can keep looping on the same best-sellers and stifle new arrivals. Set exploration rules (controlled highlighting of recent items), monitor the diversity of recommended SKUs, and align with the merchandising strategy (margin, stock, season). On the legal side, transparency about the use of personal data remains essential if you personalize based on history.
Examples by product category
Fashion and accessories
On a dress or a jacket, suggestions such as "complete the look" (shoes, bag, belt) increase the cart value when sizes and colors are consistent. For a customer who already bought a statement piece last season, a history-based logic can suggest compatible pieces or an updated range.
Electronics and accessories
Add-ons (cables, protection, warranty extensions) are natural candidates for cross-sell. Check compatibility in the attributes: an error leads to costly returns and negative reviews.
Cosmetics and parapharmacy
Multi-step routines are well suited to bundles and "next step" suggestions. Respect regulations on claims.
Food
Consumption pairings work when logistics and shelf life allow it. Respect warehouse-specific constraints if you deliver to multiple zones.
Search, collections, and synonyms
Recommendations extend discovery once the visitor is on a product page: they do not replace efficient internal search. Harmonize titles, tags, and filters with the terms your customers actually use.
Email and automation
Email journeys (abandoned cart, post-purchase) can reuse the same rules as the site if stock and prices are synchronized. See e-commerce automation.
Avoid sending three different emails with three incompatible recommendation logics: the same customer should not see one suggestion on the site, another in the abandoned cart, and a third inconsistent one post-purchase. Clear mapping of scenarios avoids the effect of "the left hand doesn’t know what the right hand is doing".
30-60-90 day roadmap
Days 1 to 30: catalog audit, blocks on high-traffic pages, basic exclusions, CTR measurement.
Days 31 to 60: extension to the cart, A/B tests, margin and returns analysis.
Days 61 to 90: homepage or email personalization for high-value segments, documentation of rules.
Best practices and common mistakes
What helps
Limit to 4 to 6 products per block on mobile.
Systematically exclude out-of-stock items and items already in the cart.
Vary labels depending on the page ("Similar" vs "Complete your selection").
Train support: they must know how to explain why a given suggestion appears.
What harms
Too many carousels on the same page.
Ignoring mobile: the main purchase button must remain visible.
Recommendations that are inconsistent with the marketing promise (e.g. a "durable" product + a disposable accessory not owned up to).
Letting a model run without review after a catalog or price change.
Forgetting to coordinate promotions: a discounted product can be suggested with a message incompatible with the discount conditions.
Checklist before validating a campaign
Displayed stock and lead times are identical between the suggestion and the product page.
Default images and variants are consistent on mobile.
User test on three journeys: new visitor, logged-in customer, add to cart then go back.
The loyalty programs and strategic promotions can be combined with targeted suggestions, but keep a margin logic.
Qstomy: recommendations in the conversational journey
Qstomy is an AI e-commerce chatbot for Shopify that can suggest relevant products while the visitor asks questions (size, compatibility, delivery). Recommendation then becomes contextual: it responds to an intent, not just a position in a carousel.
In practice, chat reduces back-and-forth between product page and FAQ: the customer gets a direct answer and a clickable suggestion in the same flow. This does not replace blocks on product pages, but complements the journey when the question is poorly phrased for classic search. For an implementation consistent with your stack, see the Shopify integration and the article chatbot for e-commerce.
Summary
Intelligent product recommendations rely on behavioral and catalog signals, often hybrid algorithms, and careful execution on product pages and cart pages. Public references on personalization (McKinsey) and AI uses on Shopify provide the business framework; concrete metrics and catalog quality determine the outcome for you. Do not promise percentage gains without measurement: test, measure, iterate.
By combining these building blocks with a well-scoped one-off promotion or a loyalty program, you align commercial incentive and product relevance rather than compensating only through price.
FAQ
Can small stores benefit from it?
Yes: content filtering and manual rules are often enough at the start; collaborative filtering improves with volume.
Do you need a data scientist?
Not to get started with a Shopify SaaS app. The main role is merchandising framing and data quality.
How do you avoid absurd suggestions?
Enrich the attributes, exclude bad associations, and monitor customer feedback on products that are often suggested.
Do recommendations replace SEO?
No: they optimize the conversion of visitors already on the site; SEO and advertising bring traffic.
Can you do without a web pixel?
You can start with the catalog and the cart; to refine behaviors, customer events are a plus documented in our complete guide to Shopify web pixels and developer documentation.
What ROI should you expect?
It depends on traffic, margin, and the quality of the suggestions. McKinsey’s work on personalization shows the performance gap between leaders and laggards: your goal is to measure your own gain, not copy a generic number.
Do you need recommendations on every page?
No: on institutional or legal pages, they distract. Focus on journeys with strong purchase intent.
How do you involve product teams?
Have a list of sensitive associations validated (exclusions, mandatory bundles, competing brands not to mix). Technology executes; merchandising sets the guardrails.
Do recommendations replace an in-store salesperson?
They mimic part of the advice by suggesting related products, but not nuanced dialogue. For stores that need advice, chat or support remain complementary.
Do you need to sync with the marketplace or POS?
If you sell across multiple channels, the availability shown in the store must reflect actual stock: otherwise recommendations lead to disappointment. Check your connectors and update delays.
How do you train internal teams?
Prepare a one-page sheet: objectives of the blocks, links to your app’s documentation, and a procedure if a suggestion displays the wrong product. The less you depend on a vendor for every micro-adjustment, the faster you iterate.
Go further
March 12, 2025
