Personalization That Drives Repeat Purchases

Product Recommendations and personalization

Part 1: How AI Learns What Your Customers Actually Want

ASOS rolled out AI-powered personalization across their site and app starting around 2018, using tools that tracked individual browsing behavior to surface different product grids for different shoppers. A customer who spent 10 minutes looking at oversized blazers would see a completely different homepage the next morning than a customer who'd been browsing activewear. Within two years, ASOS reported that personalized recommendations were contributing significantly to conversion rates, and their buyers were freed from the spreadsheet grind of manual curation to focus on trend forecasting and supplier relationships. The AI handled the 'what to show whom.' The humans focused on 'what to stock next.'

The tension ASOS faced is the same one every e-commerce team faces today: you have thousands of products and millions of possible customer combinations, but your homepage has one layout, your email has one featured item, and your team has finite hours. AI personalization doesn't eliminate the need for human judgment, it multiplies it. The principles your team sets, the customer segments you define, the brand guardrails you establish, those still come from you. The AI does the matching at scale that no human team could do manually. Understanding how that matching works is the first step to using it well.

Click to enlarge

The Core Principle: Relevance Beats Volume

Netflix discovered something counterintuitive in 2016 when they studied how customers browsed their catalog. More choice was making people less satisfied, not more. When Netflix showed users a curated row of 10 highly relevant titles, engagement went up compared to showing 40 titles with lower relevance. Their personalization team extracted a principle that applies directly to e-commerce: the right 8 products shown to the right person outperform 80 generic products shown to everyone. This is why Spotify's Discover Weekly playlist, 30 songs, perfectly matched, became one of the most engaged features in the app's history, despite the fact that Spotify has over 100 million tracks available.

The same principle shows up in e-commerce data consistently. Salesforce's 2023 Connected Shoppers Report found that product recommendations driven by AI account for 24% of orders and 26% of revenue across the e-commerce sites they studied, despite representing only 7% of site visits. That gap between traffic share and revenue share tells you everything. Customers who engage with personalized recommendations convert at a dramatically higher rate than those who browse generic category pages. They spend more. They return more often. The AI isn't creating desire, it's reducing friction between a customer and a product they were already likely to want.

For non-technical teams, the practical implication is this: you don't need to understand the mathematics behind a recommendation engine to use one effectively. What you need to understand is what signals it uses, what objectives you give it, and where human judgment still needs to override the algorithm. A beauty brand might want to suppress certain recommendations during Ramadan out of cultural sensitivity. A luxury retailer might want to ensure that sale items are never recommended alongside full-price items on the same page. These are brand decisions, not technical ones, and they're exactly the kind of input that separates a well-run AI personalization program from one that creates awkward customer moments.

Click to enlarge

A Grocery Story: How Ocado Thinks About 'Next Best Product'

Ocado, the UK's online-only grocery retailer, operates in a category where personalization logic is brutally practical. People buy food on a weekly cycle. They have strong brand loyalties. They run out of things. And critically, they don't want to spend 45 minutes rediscovering the same staples every week. Ocado's AI recommendation system does something elegantly simple: it predicts what you're about to run out of based on your purchase frequency, and surfaces those items at the top of your next shop. A customer who buys oat milk every 10 days will see it prominently at day 9. This isn't upselling, it's service.

What Ocado's approach teaches e-commerce teams outside grocery is that personalization doesn't always mean 'show them something new.' Sometimes the most valuable recommendation is a timely reminder about something they already love. For a supplements brand, that's a reorder prompt at the right moment. For a pet food retailer, it's a subscription nudge when a 5kg bag would logically be running low. The AI's job is to track the patterns that customers themselves don't consciously track, and surface the right product at the right moment, rather than the most popular product at a random moment. That shift in thinking, from popularity-based to timing-based, is one of the most underused levers in e-commerce personalization.

Comparing personalization Approaches: Manual vs. Rule-Based vs. AI

Click to enlarge

A Marketing Manager's Perspective: personalized Email at Glossier

Glossier built a cult following through community-driven marketing, but their email program faced a familiar problem by 2020: they were sending the same product-launch emails to their entire list, regardless of what individual customers had bought before. A customer who'd bought three skincare products and never touched makeup was getting the same 'new lip gloss' email as someone whose entire purchase history was color cosmetics. Open rates were decent, their brand equity carried them, but click-through rates and conversion from email were flat. The issue wasn't the creative. It was the relevance gap between what the email featured and what each recipient actually cared about.

Glossier's marketing team began using Klaviyo's AI-driven segmentation to split their email flows by purchase category affinity, essentially letting the tool identify which customers were 'skincare-first' versus 'makeup-first' versus 'fragrance-interested' based on their history. New product launch emails were then tailored to lead with the category most relevant to each segment. The result was a significant improvement in click-through rates without changing the creative quality or increasing the email frequency. The lesson for any marketing manager isn't that you need Glossier's budget or team, it's that the same Klaviyo tool is available to a 10-person brand, and the segmentation logic requires no technical expertise. You define the categories. The tool does the sorting.

Putting personalization to Work Without a Developer

The biggest misconception non-technical e-commerce teams have about AI personalization is that it requires a developer to implement or a data scientist to manage. The tools available in 2024 are built specifically for merchandisers, marketers, and e-commerce managers, not engineers. Nosto integrates with Shopify, WooCommerce, and Magento through a no-code plugin that takes under an hour to install. Once connected, it starts learning from your store's data immediately. Within two to three weeks, it has enough behavioral data to surface meaningful personalized recommendations. You configure the objectives, increase average order value, promote new arrivals, clear slow-moving stock, through a dashboard that looks like a standard marketing tool.

Klaviyo, which most e-commerce teams already use for email marketing, has built AI-powered predictive analytics directly into its platform. It can predict a customer's next purchase date, their lifetime value, and their churn risk, and it uses those predictions to trigger personalized flows automatically. If you already have Klaviyo set up for your welcome series and abandoned cart emails, you're one configuration step away from AI-personalized product recommendations in those same flows. The AI layer sits on top of what you've already built. You don't rebuild anything. You extend what's already working.

For teams on tighter budgets, Shopify's native 'Search & Discovery' app now includes AI-powered recommendation blocks that can be added to product pages, cart pages, and collection pages without any coding. It's less sophisticated than Nosto or LimeSpot, but it's free with a Shopify subscription and it works. The point is that there is no version of this conversation where 'we don't have a developer' is a valid reason not to start. The tools exist. They're affordable. They're designed for people exactly like you. The question is which touchpoint to start with and what objective to set, and those are strategic decisions, not technical ones.

Key Principles From Part 1

1. Relevance outperforms volume, eight well-matched products convert better than eighty generic ones, which is why AI personalization consistently outperforms static merchandising at scale.
2. AI personalization multiplies human judgment rather than replacing it, you set the brand rules, customer segments, and objectives; the AI does the matching across thousands of product-customer combinations.
3. The highest-value recommendations aren't always about discovery, sometimes they're about timing, like surfacing a reorder prompt exactly when a customer is likely to run out.
4. Non-technical teams can implement AI personalization without a developer, tools like Nosto, Klaviyo, LimeSpot, and Shopify's native Search & Discovery app are built for marketers and merchandisers.
5. Start with one touchpoint and measure it for 60 days before expanding, post-purchase email and returning visitor homepage are the two highest-impact starting points for most e-commerce businesses.
6. The gap between traffic share and revenue share for recommendation-driven sessions (7% of visits, 26% of revenue per Salesforce data) reveals the compounding financial case for personalization.
7. Manual curation, rule-based logic, and AI personalization are not competing choices, they work best in combination, with AI handling scale and humans enforcing brand guardrails.

When personalization Goes Wrong, and What Netflix Did About It

In 2022, Netflix faced a quiet crisis. Their recommendation engine, one of the most sophisticated in the world, was steering millions of users into what internal teams called 'content tunnels.' Subscribers who watched one crime documentary got fed an endless stream of crime documentaries. People who enjoyed a single romantic comedy were trapped in a rom-com loop. Engagement numbers looked fine on the surface, but churn was climbing. Users weren't canceling because they had nothing to watch. They were canceling because they felt like they'd seen everything the platform thought they were. The algorithm had learned their past behavior perfectly. It just couldn't imagine their future preferences.

Netflix's product team diagnosed the problem as 'over-optimization.' The recommendation system was so focused on predicting what a user would click on next that it stopped introducing variety, surprise, or discovery. It had confused short-term engagement with long-term satisfaction. Their fix involved deliberately injecting what engineers called 'exploration rows', sections of the interface designed to surface content slightly outside a user's established patterns. The goal wasn't accuracy. It was delight. This distinction, between a recommendation that is technically correct and one that actually serves the customer, is the central tension every e-commerce team faces when deploying AI personalization.

The Netflix story contains a principle that applies whether you're selling streaming subscriptions, running a boutique clothing store, or managing a B2B software catalog. personalization powered by AI is only as valuable as the outcome you're optimizing for. If you train your system purely on clicks, you get click-maximizing behavior. If you optimize for purchase completion without factoring in returns, you'll recommend products that look appealing but disappoint on arrival. The most effective AI recommendation strategies define success in human terms first, then work backwards to the data. That sequence matters enormously.

How Sephora Built a Recommendation Engine That Feels Like a Friend

Sephora's digital team spent years trying to solve a problem that sounds simple: beauty products are deeply personal. Shade matching, skin type, fragrance sensitivity, ingredient allergies, the variables are enormous. A foundation that's perfect for one customer is completely wrong for another, even if they're the same age and have the same purchase history. Traditional recommendation engines, which largely work by pattern-matching across similar customers, kept producing suggestions that were technically logical but practically useless. Recommending a moisturiser because 'people who bought your cleanser also bought this' ignores the fact that skin types vary wildly.

Their solution was to layer AI on top of explicit customer input rather than relying entirely on behavioral inference. The Sephora app's Beauty Advisor quiz gathers data about skin tone, concerns, and preferences directly from users. That self-reported data gets combined with purchase history, product reviews the user has written, and even which products they've browsed without buying. The AI then uses all of this to generate recommendations that feel genuinely tailored. The key insight from Sephora's approach: the best AI recommendations often combine what customers do with what they tell you. behavioral data alone creates a limited picture. Asking direct questions, and using AI to process the answers, dramatically improves relevance.

For non-technical professionals running smaller operations, this translates directly into practice. You don't need Sephora's engineering team to apply this principle. If you're using a tool like Klaviyo for email marketing, you can create a short post-purchase survey and use those responses to segment your recommendation emails. If you're using Shopify with an AI recommendation app like LimeSpot or Rebuy, you can configure recommendation logic to prioritize products that match attributes customers have explicitly stated they care about. The technology is accessible. What requires your input is defining which customer attributes actually matter for your specific product category.

B2B personalization: How Salesforce Proved It Works Beyond Consumer Retail

Most conversations about AI-powered product recommendations center on consumer retail, fashion, beauty, electronics, food. But the same principles apply with equal force in B2B commerce, where the purchasing cycles are longer, the order values are higher, and the cost of a wrong recommendation is far greater. Salesforce ran an internal study across their AppExchange marketplace, a platform where businesses browse and purchase software integrations, and found that AI-generated recommendations increased the rate of multi-product adoption by 27%. When companies discovered a relevant tool they hadn't been looking for, they were significantly more likely to expand their Salesforce usage overall.

What made this work in a B2B context was that the AI wasn't just matching products to individual users, it was matching products to company profiles. Industry, company size, existing tech stack, and past support ticket topics all fed into what got recommended. A 50-person marketing agency got different suggestions than a 500-person logistics company, even if both accounts had the same primary Salesforce products installed. For B2B professionals managing product catalogs, procurement platforms, or client-facing portals, this is a meaningful signal. personalization doesn't have to be about individual consumer preferences. It can be about business context, and that's often richer, more stable data to work with.

Comparing AI Recommendation Approaches: What Each Method Actually Does

Click to enlarge

The Merchandiser's Role: How a Fashion Buyer at ASOS Uses AI Without Losing Creative Control

At ASOS, merchandisers don't just accept what the algorithm serves up. They work alongside it. A senior buyer on the womenswear team described her workflow in an industry interview: every Monday morning, she reviews AI-generated reports showing which product pairings are performing above expectation in recommendation slots. She doesn't treat these as mandates. She treats them as hypotheses. If the AI is strongly pushing a particular jacket with a specific trouser, she asks why, is it because customers genuinely love the combination, or is it because both items are heavily promoted and the signal is contaminated by paid placement? That kind of critical reading of AI output is a professional skill.

Her team also uses AI to flag what she calls 'invisible inventory', products that are well-stocked, well-priced, and well-reviewed but rarely surfacing in search or recommendations because they lack the engagement history to climb the algorithm. She uses ChatGPT to write enhanced product descriptions for these items, making them richer in the kind of natural language that AI recommendation engines parse to understand what a product is actually like. Better descriptions feed better recommendations. This is a feedback loop most merchandisers don't think about: the content you put into your product listings directly shapes what the AI can do with those listings downstream.

Click to enlarge

Putting It Into Practice: Three Things You Can Change This Week

The gap between understanding AI personalization and actually deploying it tends to come down to three practical blockers: not knowing where to start, underestimating the quality of your existing data, and waiting for a 'perfect' setup before launching anything. All three are worth addressing directly. Start with what you already have. If your store is on Shopify, you already have access to recommendation features built into the platform, most merchants have them switched on but have never configured the logic or reviewed how they're performing. Spend thirty minutes in your Shopify admin looking at which recommendation sections are live on your product pages and what rules are driving them.

Second, audit your product data. Open your ten worst-performing products by conversion rate and look at their descriptions, tags, and metadata. Are they giving an AI recommendation engine enough to work with? Are the product tags consistent with how customers actually talk about those items? Tags like 'summer,' 'gift,' 'under-50,' and 'casual-friday' are the vocabulary your recommendation engine uses to connect products to moments. If your tagging is inconsistent or sparse, your recommendations will be too. This is unglamorous work, but it has a direct and measurable impact on recommendation performance, and it's something any team member can do without any technical knowledge.

Third, test one AI-generated recommendation email before building a full automation. Use ChatGPT or Claude to draft a personalized recommendation email for a specific customer segment, say, customers who bought a single product six months ago and haven't returned. Give the AI your product catalog details, the customer's purchase history, and the tone you want. Send it to a small segment manually and measure the response. This gives you real data on whether AI-assisted personalization moves the needle for your specific audience before you commit to a full platform integration. Proof of concept first. Infrastructure second.

Key Principles from Part 2

1. Define success in human terms before configuring AI recommendations, clicks, satisfaction, and long-term retention are different goals that produce different algorithmic behavior.
2. Combining explicit customer input (surveys, quizzes, stated preferences) with behavioral data produces more accurate recommendations than behavioral data alone.
3. AI personalization applies to B2B commerce as effectively as consumer retail, business context (industry, company size, tech stack) is often richer and more reliable than individual consumer signals.
4. The five main recommendation approaches (collaborative filtering, content-based, hybrid, contextual, and AI-assisted curation) suit different business types and catalog sizes, choose based on your data maturity and brand requirements.
5. Product descriptions are recommendation data, rich, detailed descriptions directly improve the quality of AI-generated matches and should be treated as a technical asset, not just marketing copy.
6. Merchandisers and marketers retain creative and strategic control over AI recommendations by treating algorithmic outputs as hypotheses to interrogate, not mandates to accept.
7. The fastest path to results is testing one AI-assisted recommendation asset, an email, a product page section, a post-purchase follow-up, before committing to full platform integrations.

When personalization Goes Wrong, and How to Get It Right

In 2012, Target's data team built a pregnancy prediction model so accurate it sent maternity coupons to a teenager before her father knew she was pregnant. The story became a cautionary tale taught in marketing schools worldwide. Target hadn't done anything technically wrong, the algorithm was working exactly as designed. What they'd missed was the human side of personalization: the moment when relevance tips into intrusion, and a customer feels watched rather than understood. The backlash forced Target to deliberately blur their targeting, mixing baby product offers with unrelated coupons to make the personalization less obvious.

This tension sits at the heart of AI-powered product recommendations. The same capability that makes a customer feel delighted, 'How did they know I'd want that?', can make another customer feel surveilled and uncomfortable. The difference is rarely in the data. It's in the judgment calls made around how, when, and how visibly that data gets used. For non-technical professionals managing e-commerce experiences, this is exactly the kind of decision that lands on your desk. You don't need to build the algorithm. You need to understand where the lines are and how to brief the people or tools working around you.

Target eventually recovered. Their personalization programs now generate hundreds of millions in attributed revenue annually. The lesson they extracted wasn't 'stop personalizing', it was 'personalize with enough context that customers feel served, not stalked.' That principle applies whether you're running a Shopify store with 500 customers or managing a product team at a mid-size retailer. The scale changes. The principle doesn't.

The Principle: Serve the Moment, Not the Profile

Spotify's Discover Weekly playlist launched in 2015 and became one of the most celebrated personalization products ever built. By 2016, users had streamed Discover Weekly tracks over 5 billion times. But the reason it worked wasn't just algorithmic precision. It was timing and framing. The playlist arrived every Monday morning, a fresh start to the week, a low-stakes moment to try something new. Spotify served the moment. They didn't surface a playlist called 'Based on your 847 listening sessions, we've profiled your exact taste.' They called it Discover Weekly and made it feel like a gift.

Netflix applies the same logic to its thumbnail personalization. The same show gets different cover images depending on who's browsing. A viewer who watches a lot of romantic dramas sees a thumbnail emphasising the love story. A viewer with a comedy history sees a thumbnail with a lighter, funnier moment from the same film. Netflix isn't changing the product, they're changing the first impression to match what that individual person is likely to respond to. This is personalization serving the moment of decision, not broadcasting a complete psychological profile back at the customer.

For e-commerce managers without a Netflix-scale engineering team, the practical translation is this: focus your personalization efforts on decision moments, the product page, the checkout upsell, the post-purchase email. These are the moments where a well-timed, contextually relevant suggestion genuinely helps a customer. A prompt to an AI tool that generates these suggestions should be anchored in what the customer just did, not everything they've ever done.

Click to enlarge

personalization in B2B: The Salesforce Playbook

personalization isn't only a consumer retail challenge. Salesforce uses AI-driven personalization across its own platform to show different feature recommendations to different types of users, a sales manager sees CRM pipeline tools highlighted, while an HR professional sees employee engagement features. This is the same principle applied to a B2B software context: understand the role, serve the relevant use case. Salesforce's Einstein recommendation engine drives this at scale, but the underlying logic, segment by job-to-be-done, not just by company size, is something any e-commerce team selling to business buyers can replicate manually with AI writing tools.

A procurement manager at a mid-size office supplies company described doing exactly this with ChatGPT. She used it to write separate product recommendation emails for three different buyer personas: office managers buying in bulk, individual professionals buying for home offices, and HR teams setting up new employee kits. Same product catalog, three completely different recommendation framings and benefit emphases. Her open rates on the segmented emails ran 34% higher than the generic version she'd been sending previously. No algorithm. Just clear thinking about who needed what, translated into copy by an AI tool.

The Content Manager's Edge: personalized Without Being Creepy

A content manager at a UK outdoor gear retailer faced a familiar problem: their newsletter was going to 40,000 subscribers with the same product recommendations every week. Engagement was flat. She didn't have the budget for a sophisticated recommendation engine, but she had ChatGPT and a basic understanding of her four main customer segments, weekend hikers, serious mountaineers, urban commuters buying weatherproof clothing, and gift buyers. She started writing four short recommendation sections per newsletter, one per segment, and using Mailchimp's basic segmentation to route each section to the right group.

Within three months, click-through rates on product recommendations had doubled. The products were the same. The prices were the same. The difference was that a weekend hiker was now reading 'Perfect for a Saturday in the Peak District' rather than a generic product description written for everyone and landing for no one. Her insight was simple: personalization doesn't require a machine learning model. It requires knowing who you're talking to and writing like you mean it. AI tools made that practical at a scale one person couldn't have managed manually.

Putting It Into Practice

The most common mistake professionals make when starting with AI-powered personalization is trying to do everything at once, full behavioral tracking, dynamic on-site content, personalized email flows, retargeting ads. The teams that get results fastest pick one decision moment and make it excellent. That might be the post-purchase email. It might be the product page cross-sell. It might be the abandoned cart message. Choose one, define two or three customer personas relevant to that moment, and use an AI tool to write recommendation copy for each. Measure the difference. Then expand.

The tools available without any technical expertise are genuinely capable. ChatGPT Plus and Claude Pro can write compelling, persona-specific recommendation copy in minutes. Klaviyo's AI features can trigger personalized email sequences based on browsing and purchase behavior without any coding. Canva AI can generate product imagery variations for different audience segments. These aren't compromises on a proper AI strategy, for most small and mid-size e-commerce operations, they are the proper AI strategy.

What separates teams that see results from those that don't isn't access to better tools. It's the quality of the brief they give the tool. A vague instruction produces generic output. A specific persona, a clear decision moment, a defined tone, and a concrete product produces something a customer actually responds to. The craft isn't technical. It's editorial. That's a skill professionals already have. AI just makes it faster and scalable.

Click to enlarge

Key Takeaways

1. personalization works when it serves the customer's current moment, not when it broadcasts how much data you've collected about them.
2. The Target pregnancy prediction case shows that algorithmic accuracy without human judgment creates customer discomfort, even when the data is correct.
3. Spotify and Netflix succeed because they frame personalization as a gift, not a profile, the same principle applies to e-commerce recommendation copy.
4. B2B personalization follows the same logic as B2C: segment by job-to-be-done, not just company size or industry.
5. Free AI tools like ChatGPT and Claude can produce persona-specific recommendation copy at scale, no technical skills required.
6. The quality of your brief to an AI tool determines the quality of the output, specificity about persona, moment, tone, and product is everything.
7. Start with one decision moment, two or three personas, and measure before expanding, teams that try to personalize everything at once rarely personalize anything well.
8. The craft of effective AI-assisted personalization is editorial, not technical, it's about knowing your customer well enough to brief a tool accurately.