Product Recommendations
Product recommendations are the "you might also like" surfaces on PDPs, carts, and post-purchase pages. Done well, they shrink decision cost and lift AOV — done badly, they add noise.
Product Recommendations
Algorithmic or curated product surfaces that suggest relevant items to a shopper on PDP, cart, or post-purchase pages.
Product recommendations are the on-site surfaces — usually labelled "you might also like", "frequently bought together", or "complete the look" — that propose additional items to a shopper based on the product they're viewing, their session behaviour, or their order history. The underlying logic ranges from manual merchandising rules to collaborative filtering and embedding-based models.
As an instance of choice architecture, a recommendation block is not just a revenue lever; it's a UX decision that shapes how a shopper navigates your catalogue. The best implementations reduce the effort of finding a relevant next item. The worst add visual noise, dilute attention away from the primary call-to-action, and slow the page down.
Three surfaces matter most. On the product detail page, recommendations help a shopper who isn't sold on the current item find a better fit. In the cart, they nudge add-ons — accessories, refills, the second item that triggers free shipping. Post-purchase, they seed the next order with complementary categories the customer hasn't bought yet.
Each surface has a different job, so each needs a different success metric. PDP recommendations are judged on click-through and assisted conversion. Cart recommendations are judged on attach rate and AOV lift. Post-purchase recommendations are judged on repeat-purchase rate at 30 and 60 days. Optimising all three against a single "recs revenue" number hides which surface is actually working.
Incremental Revenue = (AOV_with_recs − AOV_without_recs) × Orders_exposed
AOV_with_recs
AOV with recommendations exposed
Average order value for sessions where the shopper saw and could interact with the recommendation block.
AOV_without_recs
Baseline AOV
Average order value for the holdout / control group with recommendations hidden.
Orders_exposed
Orders exposed
Number of completed orders in the exposed variant during the measurement window.
A Shopify apparel brand A/B tests a 'complete the look' block on PDPs for one month.
AOV with recs: €84.20
AOV without recs (control): €78.40
Orders exposed: 4,200
→ (84.20 − 78.40) × 4,200 = €24,360 incremental revenue
A €5.80 AOV lift over a control group, applied across 4,200 orders, produced just over €24k of incremental revenue in a month. The relevant question is whether that lift is statistically significant — small AOV differences need large samples to read clean.
Two cautions when reading this formula. First, an uplift in AOV is only valuable if conversion rate doesn't drop — sometimes recommendations distract from the add-to-cart and you trade volume for basket size. Always look at revenue-per-session, not AOV in isolation. Second, holdouts have to be true holdouts (block hidden, not just empty) or you're measuring layout, not personalisation.
Typical performance of product recommendation surfaces on Shopify / Woo stores in the €1M–€15M revenue band
| Surface | Click-through rate | Revenue per session lift | AOV lift |
|---|---|---|---|
| PDP — 'You might also like' | 6–11% | +2–4% | +1–3% |
| PDP — 'Frequently bought together' | 4–8% | +3–6% | +4–8% |
| Cart — Cross-sell upsell | 8–14% | +4–7% | +5–10% |
| Cart — Free-shipping nudge bundle | 10–18% | +3–5% | +6–12% |
| Post-purchase — Thank-you page offer | 3–7% | n/a (next-order metric) | +8–15% on attached orders |
| Email — Browse / cart abandon recs | 2–5% CTR | +1–3% on flow revenue | +2–4% |
The pattern across these surfaces: cross-sells in the cart consistently outperform related-item blocks on the PDP. The shopper has already committed; you're answering "what else?" rather than "is this the right one?". Brands that under-invest in cart and post-purchase recommendations — and over-invest in PDP carousels nobody clicks — are leaving the easiest wins on the table.
Frequently asked questions
Choice architecture is the broader discipline of structuring how options are presented to influence decisions. Product recommendations are one specific application of it — the surfaces that propose which items to consider next. Recs inherit the same constraints: too many options causes paralysis, and the order of presentation shapes what gets picked.
Four to six on desktop, two to four visible on mobile with horizontal scroll. More than that and click-through rates collapse — readers default to scanning the first two items, which means anything past position three is mostly decoration. Test a tighter block before adding more slots.
They can. Each recommendation app typically adds 50–300ms of script weight and one or more network calls. If you're stacking a recs app, a reviews app, and a search app, the combined hit on Largest Contentful Paint is real. Audit the script weight before adding, and prefer apps that lazy-load below-the-fold blocks.
Both. Algorithmic recommendations win on PDPs where you have many SKUs and behavioural data is meaningful. Manual merchandising wins for collection launches, hero campaigns, and any rule the algorithm doesn't know (margin priority, end-of-season clearance). The best setups blend the two with editorial overrides on top of a personalised base.
Run a holdout test: hide the block for a randomised slice of traffic and compare revenue-per-session, not AOV alone. AOV in isolation can rise while conversion drops, leaving you flat or worse. Two to four weeks is the typical read window for a mid-traffic store.
Lightly. With no history, your best signal is the product currently being viewed, plus catalogue-wide popularity. Heavy personalisation needs identity — logged-in users, returning visitors with a meaningful session history. For anonymous first sessions, well-tuned "frequently bought together" usually beats a cold-start ML model.
Both, and they do different jobs. The thank-you page captures attention while the buying mode is still warm, often via one-click upsell apps. Email recommendations work over a longer arc — replenishment timing for consumables, complementary categories at the 30-day mark. Don't use the same product mix for each surface.
Sometimes. The strongest implementations use actual co-purchase data from your order history. Many plug-and-play apps fall back to category-based rules when co-purchase signal is thin — fine for a launching store, but worth upgrading once you have 12+ months of order data the model can learn from.
PDP-related blocks typically land at 6–11% CTR on apparel and beauty stores. Cart cross-sells often hit double digits because intent is higher. If you're below 4% on a PDP block, the issue is usually relevance (the algorithm is suggesting items too similar to the one being viewed) or placement (too far below the fold).
They can, especially when you show very similar items on the PDP — the shopper bounces from the page they were already converting on to comparison-shop. Mitigate this by leaning toward complementary items ("complete the look", accessories) above the fold, and saving similar-item alternatives for shoppers who scroll deeper.
Get an AI expert review of your site
Paste your URL — Metricuno's AI runs the same heuristic checks a senior CRO consultant would, scoring your page and prioritising the fixes that'll move conversion fastest.