Shopify Product Recommendations: The Definitive Guide

Why do product recommendations matter?

Product recommendations raise average order value: a shopper who came for one product leaves with two, because the right next item showed up at the right moment. The traffic is already paid for, so your cheapest growth is helping the shoppers who already buy, buy a little more.

Run the math as a hypothetical. Say your store sits at a $46 average order value across 300 orders a month. A relevant $14 add-on by the buy button, the strap for the watch, nudges some of those carts to $60. You bought no new visitor; you sold one more thing to shoppers already checking out. A good recommendation works on demand you have, not demand you have to buy.

The edge only shows up if the recommendation fits. A shopper is on the page for a $20 phone case, and the "you may also like" row shows them a $500 phone. The row takes up space and sells nothing, and shoppers learn to scroll past it. The enemy here is "you may also like" that does not fit: the generic row built on a thin signal that ignores what the shopper is actually doing.

So the whole game is relevance. A recommendation that fits is a salesperson; one that misses is wallpaper. The rest of this guide builds the first kind: the right type, in the right place, from real behavior, measured by what each one earns.

What are the types of product recommendations?

Sledge runs thirteen of them. Each is a different honest reason to add the next product, and the type you pick should match where the shopper is and what they are doing, not whichever sounds cleverest.

• Best sellers. What is winning across your store or a collection. Fits a homepage or a category page, where a first-time shopper has no history and "what everyone buys" is a genuinely useful signal.
• Trending. The products selling fastest over a recent window, the last seven days. Fits showing what caught on this week, which best sellers, smoothed over a longer history, can miss.
• New arrivals. Your latest drop. Fits returning shoppers who have seen the catalog and want to know what is new since last time.
• Recently viewed. The items this shopper already looked at. Fits bringing a wandering shopper back to a product they were considering, with zero guessing on your part, because they told you.
• Related products. Items connected to the one on the page. Useful, but only as good as the signal behind "related," which is the trap the next chapter covers.
• Frequently bought together. The products your shoppers genuinely buy in the same order, built from your real co-purchase data. The strongest cross-sell, because it is true for your store specifically.
• Personalized picks. Tuned to the individual shopper's behavior, so a returning buyer and a first-timer do not see the identical row.
• Collection best seller. What sells hardest inside the collection being browsed, not across the whole store. Fits a deep catalog, where the storewide list and the category list are two different answers.
• Product by collection. More from the collection the shopper is already in. Fits keeping someone inside a theme they picked themselves.
• Merchant's picks. The row you choose by hand. Fits a launch, a season, or the set you want moved this week, where your judgment beats any signal.
• Recently purchased. What other shoppers just bought. Fits proof: it sells, and it sold a moment ago.
• Most viewed. The products drawing the most attention, whether or not they convert yet. Fits spotting demand your sales numbers have not caught up with.
• Shop the last. The products down to their final units. Fits honest urgency, because the scarcity is real and your inventory says so.

Frequently bought together deserves a closer look because it is the one most often faked. The honest version sells: the camera paired with the lens shoppers actually buy with it, not a catalog-wide guess. Built from co-purchase in your own orders, it shows a shopper the thing the last hundred buyers of this item also grabbed, and that pairing sells itself. Sledge runs all thirteen recommendation types, including these seven, from one engine that feeds every placement.

Can Shopify recommend products natively?

Shopify recommends products natively in two ways: a "Featured products" block you hand-pick, and a Product Recommendations API that powers basic related and complementary products. The hand-picked block is what most stores set up in the theme editor, and it is static, so it cannot adapt to the shopper or the moment.

What it genuinely does: you choose products, and they show in a block on a page. For a curated row you want every visitor to see, a "shop our favorites" strip, it works, costs nothing, and gives you exact control over what appears. If your goal is a fixed, editorial row that does not change, native Featured products is a reasonable fit and you may not need anything more.

The limits are the flip side of "hand-picked." The same row shows to every shopper, whoever they are, whatever they are viewing, wherever they are in the journey. A first-time visitor and a returning buyer with a full history get identical suggestions, so the row fits almost no one in particular. There is no "frequently bought together" from your real orders, no "recently viewed" that reflects this shopper, no "trending" that updates as products catch on. The block cannot read behavior, so it cannot react to it.

For a manual, static job, native Featured products is fine. What it cannot do is the behavioral work, the co-purchase pairing, the recently-viewed nudge, the trending pickup, that needs an engine reading what shoppers actually do. Use native Featured products for the curated row; reach for behavior-based recommendations for everything that should adapt.

Where should recommendations go?

Recommendations go where a shopper is deciding, and the type should match the spot: frequently-bought-together by the buy button, recently-viewed and best sellers on the homepage, related products in the cart. A perfect recommendation in the wrong place still gets scrolled past, so placement is half the job.

The product page is where frequently bought together does its best work. A shopper looking at a camera is in the exact frame of mind to add the lens, so a "bought together" row built from your real co-purchase data meets them at the moment of highest intent. This is cross-sell at its most natural: not an interruption, but the obvious companion to what they are buying. Build the pairing from co-purchase, not a guess, and the row sells without a discount.

The cart is the second-highest-intent spot, and it suits the add-on. A shopper at the cart has decided to buy; one more small, related item, the batteries, the care kit, the gift wrap, is an easy yes on convenience alone. The cart drawer is where this lands, beside the free-shipping bar, so the suggestion travels with the shopper. Keep it tight and related; a cart is no place for a sprawling grid.

The homepage and collection pages suit the no-history recommendations: best sellers and trending for a first-time visitor, new arrivals for a returning one. Recently viewed works almost anywhere, because it reflects what this shopper already told you they care about. Sledge places each recommendation by surface and position, with control over layout, and one store-wide setting for whether cards carry a rating and an add-to-cart button, so the right product shows up where shoppers see it and can act without leaving the page.

How do you avoid irrelevant recommendations?

You avoid irrelevant recommendations by building them from real behavior instead of a thin signal like a shared tag, and by choosing relevance over coverage so two right offers beat six maybes. The "sneakers shown sneakers" miss and the "$500 phone next to a $20 case" miss share one root cause, and one fix.

The root cause is a shallow signal. A generic recommendation guesses from something thin, a shared tag, a matching collection, so it suggests the substitute the shopper already rejected or the upgrade ten times their budget. The sneaker shopper gets another sneaker from the "footwear" tag; the $20 case shopper gets the $500 phone from the "phones" collection. The signal is too weak to know one is a replacement for a decision already made and the other is wildly off-budget. A thin signal produces a confident, wrong suggestion.

The fix is a real signal. Build frequently bought together from co-purchase in your own orders, and the pairing is true because it already happened on your store, again and again: when people buy this, they also buy that. Build recommendations from what shoppers actually view and search this visit, and the suggestion reflects what they are doing, not a tag they share. Sledge builds recommendations from real shopper behavior, not a single shared tag, which is the difference between a row that fits and one that gets ignored. Prefer your own pairing for a specific product? Your own override beats any suggestion, product by product.

The second discipline is restraint. Even with a good engine, more recommendations is not better. A row of loosely related products under every item reads as noise, and shoppers learn to scroll past noise. Relevance beats coverage: two right offers in the spots that matter beat six maybes scattered everywhere, and the per-placement revenue line from chapter seven tells you which two are pulling their weight. When recommendations are clearly missing, the fix is the signal behind them, not more rows.

How does personalization improve recommendations?

Personalization tunes recommendations to the individual shopper's behavior, so a returning buyer and a first-timer do not see the identical row. It is the difference between "what everyone buys" and "what you, specifically, would add next," the most relevant a recommendation can get.

Start with the limit it removes. A best-sellers row is useful but blunt: it shows the same products to everyone, right for a shopper with no history and increasingly wrong for one clicking a clear trail. A shopper who keeps viewing trail shoes and searching "waterproof" is telling you exactly what they want, and a generic row ignores all of it. Personalization reads that live behavior, what the shopper clicks, adds to cart, and searches this visit, and moves the products that match to the front.

The key honesty: it works on the live session, not a profile built over months. The shopper you most want to convert is often the one who showed up two minutes ago and has never signed in, and Sledge personalizes from live session signals with no shopper login required. So it is not just for logged-in regulars; it works for the first-time, signed-out visitor whose behavior this visit is signal enough.

Two scope notes. First, personalized recommendations are one of the thirteen recommendation types, so reach for them when per-shopper relevance is the lever you want, and use best sellers, trending, or frequently-bought-together where a shared signal fits better. Second, personalized search and browse are a separate feature that reorders search results and collection pages per shopper, and both run on every plan including the entry tier.

How do you measure which recommendations sell?

You measure which recommendations sell by tracking revenue per placement, not impressions or clicks, so each spot, the product page, the cart, the homepage, reports the dollars it actually earned. The store-wide number moves for a dozen reasons; the only honest scoreboard is what each placement added itself.

Impressions and clicks lie by omission. A row can get plenty of clicks and sell nothing, or quietly drive real revenue from a spot you would have guessed was dead. Judge each placement by the money it brought in. Sledge tracks revenue per placement, so the product-page "bought together" row and the in-cart add-on each report their own number, and you read the list the way you read payroll.

This is where the relevance-over-coverage rule from chapter five gets teeth. When every placement carries its own revenue line, the loosely related rows that felt like a good idea show up earning nothing, and you cut them with evidence instead of a hunch. The scoreboard turns "I think this row works" into "this row earned this much," the only basis for keeping or killing a placement.

When two versions could work, do not argue, measure. Run the frequently-bought-together row in the spot for a stretch, then the best-sellers row in the same spot, and read revenue per placement for each; keep the one that earns more. The winner is a number, not a meeting. And if you would rather be told where to start, that is the job of Growth Intelligence: it reads your orders and points at the recommendation worth placing first, with a dollar estimate attached, then measures what it actually earned against that estimate. Every suggestion is checked against real sales, so you always know what is working. And why.