Personalized Product Recommendations
Personalized Product Recommendations are product suggestions a system shows to a specific user based on what it knows about them - past purchases, browsing behaviour, similar users’ behaviour, stated preferences, or context like season, location, and time of day. The “customers who viewed this also viewed” carousel, the personalised homepage at Amazon, the “recommended for you” row on Netflix, the “you might also like” grid at a DTC site.
Of all the different things people label “personalisation,” product recommendations are the clearest case where the approach earns its keep financially. Amazon has repeatedly cited recommendation engines as responsible for roughly 35% of its revenue. Netflix credits recommendations with retention outcomes equivalent to saving more than $1B a year in canceled subscriptions.
The core algorithmic approaches
Most real-world recommendation systems mix several of these:
Collaborative filtering. “People who liked X also liked Y.” The original Amazon and Netflix approach. Works well with dense user-item interaction data; struggles with new items and new users (the “cold start” problem).
Content-based filtering. Recommend items similar in attributes to items the user engaged with. A user who watched three historical dramas gets recommendations for other historical dramas. Works well for new items (whose attributes are known) but tends to produce narrow, predictable recommendations.
Matrix factorisation. The mathematical backbone of many modern recommenders - decomposing a sparse user-item matrix into latent factors that capture implicit taste dimensions.
Deep learning approaches. Neural networks trained on user-item interaction data, often incorporating sequence modeling (session-based RNNs, transformer-based models). Dominates current research and practice at large e-commerce and streaming platforms.
Hybrid systems. Production systems combine multiple approaches plus business-rule overrides (don’t recommend out-of-stock items, prioritise high-margin products, avoid recommending items the user already owns).
Where product recommendations succeed and fail
Succeed most clearly: E-commerce with deep catalogues (>500 SKUs) where users routinely browse rather than arrive with a specific product in mind. Media and streaming platforms with extensive content libraries. Marketplaces where discovery is the core user experience.
Fail quietly: Single-product or narrow-catalogue businesses where recommendations add noise. B2B SaaS where “product” is really one product with modules. Businesses where the underlying data is sparse (low repeat purchase rate, anonymous traffic, poor identity resolution).
Overshoot: Sites that clutter every page with four recommendation widgets, each showing six items, drawn from overlapping sources. Cognitive overload, lower conversion than simpler pages.
The privacy and signal-quality shifts
Two structural changes have reshaped what’s possible:
Third-party tracking decline. Apple’s ATT framework, Chrome’s cookie deprecation timeline, regulatory pressure in Europe. Recommendations based on cross-site behaviour are harder to build now. First-party data has become more valuable - and sites with strong login-based user identity are in a structurally better position than sites that relied on anonymous retargeting signals.
LLM-powered explanations and refinement. Large language models can now generate natural-language explanations for why an item is recommended (“This looks like what you’re shopping for because you recently viewed three stand mixers in the same price range”). The transparency can improve click-through and trust, though the research on long-run impact is still unsettled.
What “good” looks like
Four properties:
Personal, not just popular. A recommendation block that shows the site’s best-sellers to every user isn’t personalised - it’s editorial. Real personalisation looks different for different users.
Diverse within relevance. Five recommendations that are all minor variants of the same thing produce less discovery than five that cover the relevance space more broadly. Modern recommender systems explicitly optimise for diversity alongside relevance.
Updated quickly after user action. A user who just bought a coffee table shouldn’t see coffee tables recommended for the next two weeks. Fast feedback loops matter.
Respect of explicit user signals. If the user dismisses a recommendation, that dismissal is signal. Systems that keep showing dismissed items erode trust.
An example
A mid-sized furniture retailer had a “Recommended for you” row on every product page, populated by a third-party recommendation engine. The engine was configured out-of-the-box with default settings and had been running untouched for two years. Clickthrough on the recommendation row was 1.4%; attributed conversion rate from the row was 0.3%.
A three-month optimisation project made several changes. The recommendation set was scoped to the current category (e.g., on a sofa page, recommend sofas or immediate adjacencies, not random furniture). A cold-start fallback for anonymous users was added, blending category-best-sellers with in-session browsing signal. Out-of-stock items were filtered from recommendations in real-time. Items the user had already viewed in the session were down-weighted.
Clickthrough on the row grew to 4.1%. Attributed conversion rate grew to 1.7%. On a site generating $18M annual revenue, the changes corresponded to roughly $1.2M in additional attributed revenue. The recommendation technology hadn’t been the problem - it was doing what it was told. The configuration and business logic were the problem. Most off-the-shelf recommendation systems sit in this condition at most companies.
We built Penfriend to produce content that complements personalised product recommendations - category explainers, product comparison guides, use-case articles. Recommendation engines don’t work on a barren catalogue; they need content to contextualise the products being recommended.
Related terms
- Personalize - the broader personalisation concept
- Personalized Marketing - the strategic frame
- Dynamic Content - the rendering technology recommendations use
- Audience Segmentation - the strategic segmentation layer recommendations build on
- Conversion Rate - the metric recommendations are measured against