AI Product Recommendations vs Manual Picks for Upsells
AI recommendations earn their keep on long catalogs. Under 20 products, a founder who knows what sells with what usually beats the algorithm. Here's how to tell which store you are.
Here's my unpopular take: most stores shopping for AI don't have a catalog big enough to need it.
The AI product recommendations Shopify app listings promise are real on a long catalog. Hundreds of products, steady orders, a long tail nobody has time to pair by hand. There, the algorithm wins. But if you sell fewer than 20 products, a founder who knows what sells with what will usually beat the machine, because the machine is guessing from data you don't have yet.
So the answer is a split. Long catalog, let AI pick. Short catalog, pick by hand, then test.
Under 20 products, the best recommendation engine in your store is the person who packs the orders.
What does Shopify's AI actually do with your products?
Start with what you already have for free, because most founders never look under the hood.
Shopify's Search & Discovery app is built by Shopify, costs nothing, and lists "AI recommendations" as a feature. Per Shopify's own help docs, related products are generated automatically using three strategies: purchase history, product descriptions (English storefronts only), and related collections as the fallback when the first two don't have enough to go on.
Read that fallback line twice. When there's no data, it recommends from the same collection. That's the "you may also like" carousel showing three other lavender candles under a lavender candle.
And here's the part that matters for upsells. Complementary products are manual only. You choose up to 10 per product. Shopify's docs also say the purchase-history strategy needs the product to have "previous sales to assess buying behavior." No sales history, no learning.
So the free AI shows similar things. The thing that raises order value, the add-on, is a decision Shopify hands back to you.
The app is rated 2.6 out of 5 across 511 reviews on the listing as I write this. One merchant complained that a search for one product returns "100s of random unrelated products." That's what an engine looks like when it's starving for signal.
Why do manual picks win on small catalogs?
Because a small catalog is a small number of decisions, and you already know the answers.
I read a Shopify Community thread last week from a founder with "a small catalog of 15 - 20 products" that are "matching products and can be sold in bundles." Every reply pitched an app. Nobody said the obvious thing: with 18 products, you have 18 pairing decisions to make. That's an afternoon.
Picture a skincare brand with 14 products, a hypothetical. The hero is a $48 vitamin C serum. The founder knows, from 2 years of packing boxes, that serum buyers who also take the $22 SPF moisturizer come back. That pairing makes sense to a buyer in one glance: you put on the serum in the morning, the SPF goes on top.
An algorithm with 300 orders a month and 14 products sees noise. It might suggest the cleanser because both sit in the "Face" collection. It might suggest a gift card. It can't know that the SPF is the one that protects the result of the serum.
Baymard's research splits suggestions into two jobs: alternatives (help find the right product) and supplements (help finish the package). Only 42% of major sites offer both properly. The rest, 58%, offer one type or mix both in one element. Baymard notes manual curation is the most accurate approach but hard to scale. On 14 products, there's nothing to scale.
The algorithm needs thousands of orders to learn what a founder learned packing the first 200.
When does AI beat the founder?
The moment the catalog outgrows your memory.
A home goods store with 600 products can hand-pick its top 20 sellers. The other 580 get whatever the theme falls back to. That long tail is where AI earns its fee: it pairs the $19 linen napkin with the table runner people actually bought it with, on a product page you haven't opened in a year.
Two vendor guides I read this week put rough lines on this. One recommendations guide says a 50-product store doing 200 orders a month "doesn't have enough data for personalization" and suggests graduating past about 500 monthly orders. Treat that as a vendor's rule of thumb, not a law. It still points the same way I do.
Paid apps add per-visitor personalization on top. LimeSpot's listing, for example, offers AI personalized recommendations alongside manual rules, with a free plan up to 10 orders a month. Notice that even the AI apps ship a manual override. That tells you something.
Here's the whole comparison in one place.
| Manual picks | AI recommendations | |
|---|---|---|
| Best catalog size | Under ~20 products | Hundreds of products |
| Data needed | Your head and your order history | Steady sales history per product |
| New product launch | Works day one | Weak until sales build up |
| Long tail coverage | Falls apart past your top sellers | Covers every product page |
| Reason the buyer sees | "Pairs with your serum" | Often "You may also like" |
| Cost on Shopify | Free (complementary products) | Free related products, paid apps for personalization |
| Biggest failure | Stale picks nobody revisits | Similar items instead of add-ons |
What does the add-to-cart gap look like?
I went looking for a published study comparing add-to-cart rates on hand-picked versus algorithmic suggestions. I couldn't find one. So the chart below is a hypothetical illustration of the pattern I'd expect, not measured data. Don't quote it as a benchmark.
Illustration only, not measured data. Rates are percent of product page visitors who add the suggested item.
The published data that does exist is about recommendations in general. Salesforce data on 150 million shopping sessions, as compiled in this recommendation statistics roundup, found that visits with a recommendation click were 7% of visits but drove 26% of revenue. Read that carefully. It counts people who chose to click, and clickers are already deep in buying mode. It proves recommendations sit in the path of big buyers. It doesn't prove the algorithm beat a person.
How much is one good manual pick worth?
Here's the math on that hypothetical skincare store.
Before: conversion rate 2.1%, average order value $48. That means revenue per visitor is $1.01. On 10,000 visitors, that's 210 orders and $10,080.
Now 36% of buyers add the $22 SPF. That's $7.92 more per order on average, which lifts the average order to roughly $56. Conversion rate holds at 2.1%, average order value $56. Revenue per visitor: $1.18. Same 10,000 visitors, same 210 orders: $11,760.
That's $1,680 a month from one checkbox in Search & Discovery. No app. No model. A founder deciding.
And here's where I'll admit something. I used to treat order value as a widget problem. Install the app, wait for the AI to learn. It took me too long to see that the biggest order-value jumps I've watched came from the page itself making a bigger order the obvious choice.
One of the Amazon-to-Shopify products on our results page went from a $29.95 average order on Amazon to $177.75 on Shopify, and from $7.73 to $78.87 net profit per sale after ads and inventory. Real client numbers, not typical results, and not a promise of what your store will do. That order size came from an offer built into the page on purpose, not from a carousel guessing.
How do you pick the right one for your store?
Ask three questions. Answer honestly.
How many products do you sell? Under 20, start manual. Between 20 and roughly 100, hand-pick your top 10 sellers and let Shopify's free related products cover the rest. Hundreds, AI starts paying.
How much order history does each product have? A product with no sales gives the purchase-history strategy nothing to learn from. That's every new launch, which is often your highest-margin product. Hand-pick for launches, always.
What job is the slot doing? If it's "finish the package," that's a supplement, and you probably know it better than any model. If it's "help me find the right size or style," alternatives across a large catalog are where AI is strong.
Then test it. Two weeks manual, two weeks AI, same slot, same label, top three product pages. The scoreboard is revenue per visitor on those pages. Clicks on the widget will lie to you, because clickers were buying anyway.
Where the pick shows up matters as much as who picked it. A single supplement in the cart is one of the cheapest wins in ecommerce, and these cart drawer upsell practices cover the placement. If the buyer is already committed, the post-purchase upsell playbook is the slot where a hand-picked add-on can't cost you the first sale. And if your pairs are obvious enough, they're a bundle waiting to happen; these Shopify bundle ideas show how. The whole order-value picture lives in the average order value playbook.
Frequently asked questions
Are AI product recommendations on Shopify worth it for a small store? Usually not yet. Shopify's docs say purchase-history recommendations need previous sales, so a store with 15 products and thin order history gets guesses. Hand-pick first.
Does Shopify have free AI product recommendations? Yes. Search & Discovery generates related products automatically. Complementary products, the add-ons, are manual only, up to 10 per product.
When do AI upsell recommendations beat manual picks? When the catalog is too big to hand-pick and you have the order volume to feed the model. Hundreds of products and a long tail you'll never personally pair.
How do I test manual vs AI cross sell recommendations? Same slot, same label, two weeks each, top three product pages. Compare revenue per visitor on those pages, not widget clicks.
What should a manual product recommendation be? The one item a buyer of this product needs next. One clear pick with a reason beats ten loosely related products.
What to do next
Open Search & Discovery today and add one complementary product to your best seller. One. The item you'd hand a customer at a market stall if they asked what goes with it. Then write down your revenue per visitor on that page and check it again in 14 days.
Book Your Profit Audit
A recommendation, manual or AI, can only add to an order the page already closed. On a profit audit we'll show you where your page is leaking buyers before the add-on ever loads, and how to rebuild a high-converting product sales page in less than 15 minutes.
Or go here to check it out → revenueflows.ai
P.S. The algorithm guesses from your data. You know from your customers. On a short catalog, trust the one who's packed the boxes.
Frequently asked questions
Are AI product recommendations on Shopify worth it for a small store?
Usually not yet. Shopify's own docs say purchase-history recommendations need previous sales to work, so a store with 15 products and thin order history gets guesses. Under about 20 products, hand-pick the one item that finishes each hero product and test that first.
Does Shopify have free AI product recommendations?
Yes. Shopify's free Search & Discovery app generates related products automatically from purchase history, product descriptions (English storefronts only) and related collections. Complementary products, the add-on items, are manual only, up to 10 per product.
When do AI upsell recommendations beat manual picks?
When the catalog is too big to hand-pick and you have the order volume to feed the model. Hundreds of products, steady daily orders, and a long tail of items you'll never personally pair. That's where an algorithm covers ground a founder can't.
How do I test manual vs AI cross sell recommendations?
Run both on your top three product pages for two weeks each, same slot, same label, and compare revenue per visitor on those pages. Not clicks on the widget. A store converting 2.1% at a $48 average order value earns $1.01 per visitor; the winner is whichever setup moves that number.
What should a manual product recommendation be?
The one item a buyer of this product needs next: the refill, the case, the matching pillowcase. One clear pick with a reason in the label beats ten loosely related products, because the buyer can say yes without starting a new decision.