AI Product Recommendations on Shopify: What Actually Drives Revenue

Why generic recommendations underperform

The "customers also bought" widget shows the same products to every visitor regardless of who they are. A first-time visitor browsing casually sees the same recommendations as a customer who's bought three times and has strong category preferences.

The logic: product A and product B are frequently purchased together, so recommend B to everyone who views A. This is correlation-based, not behaviour-based — it tells you nothing about whether the specific customer currently viewing product A has any interest in product B.

AI-driven recommendations build a different model: this specific customer, or this type of customer if you don't have purchase history, has browsed category X, bought product Y 60 days ago, has a known AOV, and currently has product Z in their cart. Given all of that context, what is the highest-probability next purchase?

The difference in outcome is consistent: we typically see 15–25% higher AOV from AI-driven cross-sell vs. generic co-purchase recommendations.

The recommendation placements that actually convert

Not all recommendation placements are equal. In order of conversion impact:

• Post-add-to-cart modal: after a customer adds to cart, show 2–3 "complete the look" or "often bought with" recommendations before they continue shopping. This is the highest-intent moment on the page — the customer has already committed to buying once. Conversion rate from this placement: 8–15%.

• Cart page upsell: below the cart contents, 2–3 curated items that complement what's in the cart. Works best when the recommendation logic is transparent ("People who bought this also loved these").

• PDP cross-sell section: below the product description, after reviews. The customer is reading the page carefully — if they're research-mode, a relevant recommendation can expand the consideration set.

• Post-purchase upsell page: shown after payment, before the order confirmation. The customer has already bought — their guard is down. Conversion rates of 3–8% on post-purchase upsells are common for well-targeted recommendations. Shopify's Checkout Extensibility enables this without a third-party app.

Implementation options: from simple to sophisticated

Three levels of AI recommendation implementation, in order of sophistication and investment:

Level 1 — Enhanced native (cheapest): an app like LimeSpot or Product Personaliser that improves on Shopify's native recommendations with basic machine learning. Setup takes 1–2 days; lift vs. native is typically 10–20%.

Level 2 — Behaviour-based platform: apps like Rebuy, Recombee, or Nosto that build user-level behavioural models. Setup takes 1–2 weeks; lift vs. generic recommendations is typically 15–30% on AOV.

Level 3 — Custom recommendation engine: a model trained specifically on your catalogue and customer data using a custom ML pipeline. Makes sense for brands with 50,000+ customers and complex catalogue relationships. Setup takes 6–10 weeks, with the highest ceiling for performance and the most control over the recommendation logic.

The data you need before any of this works

Recommendation quality is bounded by data quality, and most disappointing implementations trace back to the inputs rather than the algorithm.

Cold start is real. A behavioural model needs behaviour. A store with a few hundred orders doesn't have enough signal for user-level personalization to beat a well-curated manual rule, and paying for a Level 2 platform at that stage usually buys a worse experience than hand-picking three complementary products per collection. The rough threshold where behavioural models start meaningfully outperforming curation is a few thousand orders with repeat-purchase patterns in them.

Product data has to be structured. A model can only find "complementary" relationships if your catalogue encodes what products actually are — category, material, colour family, size system, use case. Stores where all of that lives in unstructured description text get noticeably worse recommendations than stores using metafields and metaobjects properly.

Inventory awareness is non-negotiable. Recommending out-of-stock products is worse than showing nothing — it's a dead end at the exact moment you had the customer's attention. Confirm any app or engine you're evaluating filters on real-time availability, not a nightly sync.

Measuring whether recommendations are actually incremental

Every recommendation app reports "revenue from recommendations," and that number is almost always inflated — it counts any order where a customer clicked a recommended product, including customers who would have found and bought that product anyway through normal browsing.

The honest measurement is a holdout test: serve recommendations to 90% of traffic, withhold them entirely from 10%, and compare AOV and conversion rate between the two groups over a few weeks. Most platforms at Level 2 and above support this natively. The gap between the groups is your genuine incremental lift — and it's routinely a fraction of the attributed number the dashboard shows.

This matters practically because recommendation apps are usually priced as a percentage of attributed revenue or a tier based on order volume. Paying a revenue share on inflated attribution is a real cost, and knowing the true incremental number is what tells you whether upgrading from Level 1 to Level 2 is actually worth the price difference.

FAQs

How much AOV lift can I realistically expect from AI recommendations? 15–25% higher AOV vs. generic co-purchase recommendations is the typical range we see, scaling further with more sophisticated behaviour-based or custom implementations.

Where should I place my first AI recommendation if I can only add one? The post-add-to-cart modal — it's the highest-intent moment on the page, since the customer has already committed to buying once, and it converts at 8–15%.

Do I need a custom-built engine, or is an app enough? An app is enough for most brands — custom engines only make sense once you have 50,000+ customers and catalogue relationships complex enough that off-the-shelf behavioural models leave real revenue on the table.

Why do post-purchase upsells convert well despite coming after payment? The customer's guard is down immediately after committing to a purchase, and a well-targeted, relevant offer at that moment converts at 3–8% — a meaningful incremental lift with zero additional acquisition cost.

How many orders do I need before behavioural recommendations beat manual curation? Roughly a few thousand, with real repeat-purchase patterns. Below that, hand-picked complementary products per collection usually outperform a behavioural model with insufficient signal.

Is the "revenue from recommendations" number in my app dashboard accurate? No — it over-attributes by counting customers who would have found those products anyway. Run a holdout test (90% served, 10% withheld) to measure genuine incremental lift, which matters directly if you're paying a revenue share.

Our take

The gap between generic and AI-driven recommendations isn't marginal — it's the difference between a widget nobody notices and a system that meaningfully lifts AOV on every order. Start at whichever level matches your customer base size and complexity; don't jump to a custom build before an off-the-shelf platform has proven the concept, and run a holdout test before assuming the dashboard's attributed revenue is real. If your store still runs Shopify's native "also bought" widget, talk to us — even Level 1 is usually a quick win.