Recommendation Systems
Recommendation systems power the "you might also like" carousel and the "frequently bought together" upsell. Here's how the three algorithm families work, where they pay off on a product page, and the uplift you can realistically expect.
Recommendation Systems
Algorithms that surface relevant products to a shopper based on behaviour, catalogue similarity, or popularity signals.
A recommendation system is the engine behind the 'you might also like' carousel on a product page, the 'frequently bought together' bundle in cart, and the personalised module on the homepage. It scores every product in your catalogue against the current shopper context — what they viewed, who they look like, what's trending — and returns a ranked shortlist.
Three algorithm families dominate online retail: collaborative filtering (people-like-you), content-based matching (products-like-this), and popularity-based ranking (what's hot right now). Most production stacks blend all three, weighted by where the widget sits in the funnel and how much behavioural data the visitor has generated.
Collaborative filtering learns from co-behaviour: shoppers who viewed this dress also viewed these three. It needs traffic volume to work — typically 10k+ monthly sessions before the matrix is dense enough to recommend confidently. Below that threshold, it cold-starts badly and recommends bestsellers to everyone.
Content-based matching reads product attributes — category, colour, fabric, price band, brand — and recommends items that share them. It works from day one on a thin catalogue but plateaus quickly: a shopper looking at a black tee gets shown ten more black tees, not the belt that completes the outfit. Popularity ranking is the fallback for anonymous visitors and new SKUs with no signal yet.
Recommendation Lift = (AOV_with_recs - AOV_without_recs) / AOV_without_recs
AOV_with_recs
Average order value (recommendation cohort)
Mean order value for sessions that interacted with a recommendation widget.
AOV_without_recs
Average order value (control)
Mean order value for sessions where the widget was hidden or not engaged with.
A Shopify apparel store A/B tests a 'frequently bought together' module on the cart page for one month.
AOV with recs (treatment): €84.00
AOV without recs (control): €72.00
→ 16.7% lift in AOV
A 16.7% AOV lift on the cart-page bundle is in the strong band for apparel — most stores see 8-15% from this placement. Pair it with a small bundle discount and the lift typically rises another 3-5 points.
The measurement trap most teams fall into is attributing too much. If your widget recommends a bestseller, some of that revenue would have happened anyway. Always measure incrementality with a hold-out group rather than counting any click-through-then-purchase as 'recommendation revenue' — the latter inflates impact by 2-3x.
Typical revenue uplift by recommendation placement (online retail, blended verticals)
| Placement | AOV uplift | Revenue per visitor uplift | Notes |
|---|---|---|---|
| Homepage 'trending now' | 1-3% | 2-4% | Mostly popularity-ranked; thin personalization signal. |
| PDP 'you might also like' | 3-7% | 5-10% | Content-based matching dominates; relevance matters more than novelty. |
| Cart 'frequently bought together' | 8-15% | 6-12% | Highest-intent placement; pair with a bundle nudge. |
| Post-purchase upsell | 5-12% | 4-9% | One-click add; works best with accessories under 30% of AOV. |
| Email 'we picked these for you' | 10-20% | n/a (channel-specific) | Click-through quality is the binding constraint. |
Recommendation systems are one branch of personalization — the part that ranks products. The other branches (content, offers, segment-targeted creative) sit alongside. On Shopify and WooCommerce, the lightest path to production is a tag-based widget that reads the current product's category and price band; the heaviest is a real-time collaborative filter trained nightly on your full event stream. Most stores in the €1M-€15M range get 80% of the upside from the former and never need the latter.
Recommendation systems FAQ
Collaborative filtering looks at shopper behaviour ('people who viewed this also viewed…'). Content-based matching looks at product attributes ('items in the same category, price band, and colour family'). Collaborative filtering needs traffic volume; content-based works from day one.
Content-based and popularity-ranked recommendations work at any traffic level. Collaborative filtering typically needs at least 10,000 monthly sessions and a few thousand purchases per month before the co-purchase matrix is dense enough to outperform a simple bestseller list.
Cart-page 'frequently bought together' bundles deliver the largest AOV uplift (typically 8-15%) because intent is already high. Product detail page carousels drive discovery but smaller per-session lift. Homepage recommendations are mostly a merchandising signal, not a revenue lever.
The cleanest measurement is an A/B test with a hold-out group: half of sessions see the widget, half don't. Compare AOV, revenue per visitor, and conversion rate between cohorts. Avoid 'attributed revenue' counts — they overstate impact because some purchases would have happened anyway.
A well-built widget adds 30-80ms of render time and 50-150KB of JavaScript. Poorly built ones can add 500ms+ and block largest contentful paint. Test before and after with PageSpeed Insights or Lighthouse, and pin the script to load after first paint.
For stores under €5M revenue, a well-rated Shopify app (LimeSpot, Rebuy, Boost AI) covers 95% of the upside with no engineering cost. Building in-house only pays off when you have a unique catalogue structure (configurable products, B2B price tiers) that off-the-shelf apps mis-rank.
Recommendation systems are one component of broader personalization. Personalization also covers content blocks, offers, email cadence, and segment-targeted creative. Recommendations are the product-ranking piece; the rest is delivered through CMS rules, email platforms like Klaviyo, and segment-based experimentation.
This is the cold-start problem. When the algorithm lacks behavioural signal — small catalogue, low traffic, anonymous visitor — it falls back to popularity ranking. Fix it by enriching product attributes (so content-based matching has something to work with) and by capturing first-visit signals like category clicks.
It's a cart-stage recommendation showing 2-4 items historically purchased alongside the current basket — usually accessories or consumables under 30% of the main item's price. It works best on apparel, beauty, and electronics where complementary SKUs are obvious. Pair it with a small bundle discount to push another 3-5 points of attach rate.
Yes. Collaborative filtering uses your first-party event stream (session ID, viewed products, cart contents), none of which depend on third-party cookies. Content-based matching needs no shopper data at all — only the catalogue. The cookie deprecation mostly affects cross-site ad retargeting, not on-site recommendations.
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.