
On this page
- The three types of AI product recommendations
- What Shopify gives you natively: Search & Discovery and the recommendations API
- Where Shopify’s native recommendations run out
- Native vs apps vs custom: picking the engine
- Where recommendations belong: PDP, cart, post-purchase and email
- The merchandising rules that matter more than the model
- Do AI product recommendations really increase revenue?
- How to measure lift with a holdout test
- Frequently asked questions
- What we’d do first
AI product recommendations on Shopify work when three things are true: the widget sits where the shopper is already deciding, the products it shows are in stock and priced near the item in view, and you can prove the lift with a holdout group. The algorithm matters less than vendors admit. Shopify’s free engine handles “you may also like” well enough for many stores, paid apps earn their fee in the cart and email, and custom builds suit a small group of brands.
The question we hear most comes from teams who installed an app a year ago, see a dashboard full of “influenced revenue”, and can’t tell whether any of it is new money. Good question. Below: the recommendation types, what Shopify gives you natively as of November 2026, when an app or custom build pays, where each widget belongs, and how to run a holdout test that tells you the truth.
Key takeaways
- There are three jobs: complementary (“pair it with”), related (“you may also like”) and personalized (based on the shopper). Complementary is the one that raises order value.
- Shopify’s native recommendations are product-to-product, not personalized, and ignore orders imported from another platform.
- Pay for an app once you want cart, post-purchase and email placements with rules and testing. Rebuy’s cart package starts at $45 a month.
- “Influenced revenue” is not lift. In one randomized experiment, related recommendations raised combined sales by about 11% while taking some sales from the product on the page.
- Small stores can’t measure small lifts. At 60,000 sessions a month, a 50/50 test needs about 2.7 months to detect a 10% conversion lift.
The three types of AI product recommendations
Every recommendation widget does one of three jobs, and mixing them up is the most common reason one underperforms. Pick the job first, then the tool.
Complementary: frequently bought together
Complementary products are add-ons: the case for the phone, the socks for the shoes, the refill for the starter kit. This is the widget that raises average order value, because the shopper keeps the item they came for and adds a second one. In Shopify’s help center, the guidance is to pick add-ons that cost the same as the main product or less. We agree: a $14 add-on next to a $60 item is a small yes, a $90 one is a second decision.
Related: similar items
Related products are substitutes. Someone on a navy linen shirt sees two other linen shirts. It rescues sessions where the first product is wrong (size, color, price), so it lifts conversion more than order value. It also competes with the product on the page, which matters when you measure it.
Personalized: based on the shopper
Personalized product recommendations use who the visitor is, not just what they’re looking at: products they viewed, carted or bought, plus what similar shoppers did. This is where the “AI” in the marketing is most real, and where you need traffic and history before the model has anything useful to say. A store with 5,000 sessions a month, mostly first-time visitors, gets little here that a bestseller rail wouldn’t do.
What Shopify gives you natively: Search & Discovery and the recommendations API
Shopify product recommendations come in two pieces at no extra cost: automatic related recommendations for every product, and the Shopify Search & Discovery app, which is free and lets you override them by hand. Most Online Store 2.0 themes already ship a “related products” section that calls the engine.
The engine underneath is the Product Recommendations API. Per the Ajax API reference, it takes a product_id, a limit of 1 to 10, and an intent of either related (the default) or complementary. Unpublished products return a 404. Note what’s missing: there is no customer parameter. Native recommendations are item-to-item, so two very different shoppers on the same page see the same list.
For related products, Shopify works down a fallback chain. It tries purchase history first, then product descriptions, then related collections. Two details catch out a lot of merchants:
- The description strategy only works on English storefronts. A German store with thin sales history skips to collections.
- Imported orders don’t count. Shopify’s theme developer guide says orders imported from other platforms are not used. Migrated from Magento last quarter? The engine is starting from zero.

Setting manual picks in Search & Discovery
Go to Apps > Search & Discovery > Product recommendations > Add recommendations, choose a product, and add up to 10 complementary and up to 10 related products. For related, show only your picks or your picks ahead of automatic ones. Keep the automatic ones on, so the block still fills when a manual pick sells out. The bulk editor on the same page saves hours across a full collection.
Complementary products need two more things: your theme needs a complementary products block on the product template (Online Store > Edit theme > Products > Default product > Add block), and every pick needs stock above zero. Shopify allows one complementary block per product page.
From our team: the Search & Discovery listing advertises AI recommendations for both related and complementary products, but the help center describes complementary products as the ones you pick. Plan as if complementary means curation. Hand-pick add-ons for your top 50 products by revenue. That covers most of the traffic that will see the block, and it takes an afternoon.
Where Shopify’s native recommendations run out
Native AI product recommendations run out at four points: personalization, placement, rules and testing. If none of those matter yet, stay native. Plenty of stores under about $1M a year do fine with automatic related products plus curated add-ons.
- No personalization. Because the API takes a product, not a person, a returning customer who bought a starter kit last month sees the same rail as a first-time visitor.
- Placement is mostly product pages. The API needs a
product_id. You can call it from a cart drawer using the last item added, and we do, but that’s theme JavaScript, not a setting. - Few business rules. The developer guide says you can’t make the algorithm exclude products; you filter results after they come back. Sold-out items, $0 products, gift cards and cart items are dropped for you. Margin and brand rules are not.
- No built-in A/B testing. The app reports click-through and conversion for recommendations, which is useful. It can’t tell you what would have happened without them.
Native vs apps vs custom: picking the engine
For AI product recommendations, start native, move to an app when you need cart, post-purchase or email placements with rules and testing, and go custom only when no app can model your catalog logic.
| Shopify native | Recommendation app | Custom build | |
|---|---|---|---|
| Cost | Free | Monthly fee, usually scaled by orders (Rebuy’s cart package: $45 at 500 orders, $425 at 5,000) | Our planning range: 15–40 hours for theme-level work on the native API, 150–400 hours for your own model and service |
| Personalized to the shopper | No | Usually yes, using the vendor’s model | Yes, on your own data and rules |
| Placements | Product page, plus cart with custom JS | PDP, cart drawer, checkout, post-purchase, sometimes email | Anywhere you build it, including apps and headless storefronts |
| Business rules | Manual picks, stock and cart exclusions | Rules engines (margin, tags, inventory, segments) | Whatever you can specify |
| A/B testing | None | Often included or a paid add-on | You build or wire it in |
| Best for | Most stores under ~$1M a year | Growing stores with repeat buyers | Large catalogs with compatibility logic, bundles or B2B rules |
When a frequently bought together app pays for itself
Do the arithmetic before the trial ends. Rebuy’s pricing page charges per package by average monthly orders: the Cart & Merchandising package is $45 a month at up to 500 orders and $425 at 5,000, with a 14-day trial. Say, hypothetically, you do 2,000 orders a month at a $65 average order. The cart package lists at $176 a month. If the app adds $1.50 of genuinely new revenue per order, that’s $3,000 a month in sales, well clear of the fee once you apply your margin. If the true lift is $0.20, it isn’t. Only a holdout tells you which.
Every app also adds script weight, sometimes enough to become the heaviest thing on the page. See our guide to auditing the apps slowing down your store.
When custom is the right call
Custom makes sense when recommendations depend on facts a model can’t learn from clicks: which filter fits which machine, which part fits which model year, which wholesale buyer may see which SKUs. There we build the logic on metafields and order data and serve it through the theme or a Hydrogen storefront. If that’s you, our shopify ai solutions team scopes these as a short discovery first, so you find out whether an app gets you 80% of the way before anyone writes a model.
If your question is really about a chat assistant that answers “which one should I buy?”, that’s a different tool with different risks. We cover it in AI shopping assistants on Shopify.
Where recommendations belong: PDP, cart, post-purchase and email
Put complementary AI product recommendations where the shopper has already decided (cart, post-purchase, order emails) and related recommendations where they’re still deciding (the product page). Weak setups we audit often have this backwards: a “you may also like” rail in the cart, inviting a decided shopper to reconsider.

Product page
Two blocks, in this order: a small complementary block near the add-to-cart button, then a related rail lower down. Shopify’s developer guide suggests showing four products even though the API returns up to ten, and that matches what we see. Past four, the rail becomes a second collection page. The page itself matters more than the widget; see our Shopify PDP teardown.
Cart drawer
The cart is the best place for add-ons, and the hardest natively. Pass the most recently added product to the API with intent=complementary, render one to three items with a one-tap add button, and hide the block when the cart already contains the add-on. A free-shipping progress bar next to a low-priced add-on works because it gives the shopper a reason, not just a product.
Post-purchase and thank-you page
A post-purchase offer appears after payment, so it can’t hurt conversion on the first order. It’s the lowest-risk placement on the store. One product, ideally a consumable or accessory that ships in the same box.
Email is where personalization works on almost any store, because you know who the recipient is. Klaviyo’s product feed docs describe recently viewed, added-to-cart and an AI “may also like” feed built on collaborative filtering. The model takes 2 days to a week to train, retrains every 2 to 7 days, and falls back to a non-personalized default for customers with little history. One setting to check: Added to Cart tracking has to be switched on manually for Shopify.
Not sure your recommendations are adding anything?
Send us your store URL and the apps you run, and we’ll map every recommendation placement, its type and its script weight, and propose a holdout test sized to your traffic.
Book a recommendations audit
The merchandising rules that matter more than the model
With AI product recommendations, rules decide more of the outcome than the algorithm does. A great model showing an overpriced or low-margin product still loses. The rules we set first:
- Stock floor. Related products can still appear when sold out if “continue selling when out of stock” is on. Decide if you want that.
- Price band. Add-ons at or below the main product’s price, as Shopify’s help center suggests. For related items, stay within roughly the same price tier as the product in view.
- Variant sanity. If you sell combined listings, choose in Search & Discovery > Settings whether recommendations show the parent, the child or both. Five colorways of one jacket as “related” wastes the rail.
- Data quality. Product type, tags and metafields feed every engine. Fix them before paying for a smarter model.
Watch out: recommendation apps often discount the add-on automatically to “close” it. A 15% discount on every cross-sell turns an AOV gain into a margin loss. Run the numbers on contribution margin, not revenue, before you switch discounts on.
Do AI product recommendations really increase revenue?
Yes, usually, but by less than vendor dashboards suggest, and partly by moving sales around rather than creating them. The best evidence is from randomized field experiments.
In Kumar and Hosanagar’s nine-week experiment at a mid-sized US fashion retailer, showing up to four substitute products raised combined sales of the product and its recommended items by about 11%. The same study found sales of the product on the page dropped by about 1.9% once a shopper had viewed it, because some buyers switched to a recommended item. Gains were uneven: well-connected products gained a lot, the least-connected almost nothing.
A 2026 study in MIS Quarterly by Sun and colleagues adds a twist. There, recommendations lifted recommended products but cut sales of others, so the short-run net effect on total sales may not be positive at all. Long run, they improved loyalty, and switching them off cut later sales because shoppers came back less.
So “revenue from recommendation clicks” overstates the gain, and a two-week test can understate it. Both point to a holdout that runs long enough.
How to measure lift with a holdout test
A holdout test randomly hides AI product recommendations from a fixed share of visitors for the whole test, then compares revenue per visitor between groups. It’s the only measurement that answers “what would have happened without this?” Our setup on a standard theme:
- Bucket visitors in the browser. Shopify caches storefront pages, so assign the bucket client-side: a random number stored in a first-party cookie, set once and kept for the whole test. Native recommendation sections already load by JavaScript, so hiding them is easy.
- Write the bucket onto the order. Add it as a cart attribute through
/cart/update.jsso every order carries “reco_test: control” or “reco_test: on”. It survives checkout, so you analyze orders, not pixels. - Pick one primary metric. Revenue per visitor, because it captures conversion and order value. AOV and conversion rate are secondary.
- Size the test before you start. Decide the smallest lift worth detecting and the sessions it takes (below). Don’t stop early because the graph looks good.
- Keep everything else still. No big promotions, theme releases or price changes mid-test.
With an app, check it offers a true holdout (recommendations hidden entirely), not just widget-versus-widget tests. Either way, cross-check against your own order data, as in our Shopify GA4 tracking guide.
Worked example: how long your test needs to run
Take a hypothetical store with 60,000 sessions a month converting at 2.0%. Using the standard two-proportion sample size formula (95% confidence, 80% power), this is the total traffic needed to detect a relative lift in conversion:
| Lift to detect | 50/50 split: sessions | 50/50: months at 60k | 90/10 holdout: sessions | 90/10: months at 60k |
|---|---|---|---|---|
| 5% (2.0% to 2.1%) | ~630,000 | 10.5 | ~1,750,000 | 29.2 |
| 10% (2.0% to 2.2%) | ~161,000 | 2.7 | ~448,000 | 7.5 |
| 20% (2.0% to 2.4%) | ~42,000 | 0.7 | ~117,000 | 2.0 |

The uncomfortable part: a whole-store lift from recommendations is more likely to be in single digits than 20%. A store this size can’t prove a small effect in reasonable time, and revenue per visitor is noisier still. At this traffic we’d run a 50/50 split for six to eight weeks, look for 10% or more, and read “no detectable difference” as a small win at best. At 500,000 sessions a month, a permanent 5% to 10% holdout becomes realistic, and it keeps vendors honest at renewal.
From our team: a common shortcut is to compare shoppers who clicked a recommendation with shoppers who didn’t. Skip it. Clickers are more engaged to begin with and would convert better even if the widget were blank.
Frequently asked questions
Do product recommendations increase AOV?
Complementary recommendations usually do, because the shopper keeps the main item and adds a cheaper one. Related recommendations mainly lift conversion and can even lower AOV by steering buyers to a cheaper substitute. In one randomized experiment, related recommendations raised combined sales by about 11% while cutting the focal product’s sales slightly. Test AOV and revenue per visitor with a holdout rather than trusting an app’s influenced-revenue figure.
Best way to add AI recommendations to Shopify?
Start with Shopify’s built-in recommendations and the free Search & Discovery app, then add a paid app only when you need cart, post-purchase or email placements with rules and testing. Curate complementary picks for your top 50 products, keep automatic related products on, and run a holdout before you renew any paid plan. Go custom only for compatibility or B2B logic no app can model.
Does Shopify have built-in product recommendations?
Yes, Shopify generates related product recommendations automatically for every product and lets you add manual picks in the free Search & Discovery app. The engine uses purchase history, then product descriptions on English storefronts, then related collections. It is product-to-product rather than personalized to each shopper, and it does not use orders imported from another platform.
How many product recommendations should I show?
Show about four products in a related rail and one to three in a cart or complementary block. Shopify’s API returns up to ten, but its own developer guide suggests displaying four. Beyond that, the widget becomes a second collection page and pulls attention from the add-to-cart button.
What is the difference between related and complementary products on Shopify?
Related products are substitutes for the item in view, while complementary products are add-ons bought alongside it. Shopify generates related products automatically and shows them under headings like “You may also like”. Complementary products are picked in Search & Discovery, need stock above zero, and appear in a theme block such as “Pair it with”. Related lifts conversion, complementary lifts order value.
What we’d do first
Before you buy anything, open Apps > Search & Discovery and see what your store already does. Then, in order:
- Audit placements. List every recommendation widget, its type (related, complementary, personalized) and where it sits. Move any “you may also like” out of the cart.
- Curate the top 50. Hand-pick complementary add-ons for the products that bring in most of your revenue, priced at or below the main item.
- Fix the inputs. Clean product types, tags and metafields. If you migrated recently, expect native recommendations to be weak for a few months while order history builds up.
- Add one paid placement, not five. If you go to an app, start with the cart drawer, where complementary items do the most work.
- Hold out and measure. Size the test to your traffic, write the bucket to the order, and judge the app on revenue per visitor at renewal.
Most stores that do this end up with fewer widgets that each do more. We’ve built for brands like Delsey, EcoFlow and Kerbholz. If your catalog needs logic an app can’t express, or you want the holdout set up right the first time, talk to our Shopify AI and personalization team.




