E-commerce
March 12, 2025
Do you want to offer truly relevant suggestions without drowning your customers in generic blocks? Purchase history is one of the most reliable data points for personalizing recommendations: a purchase confirms intent, a budget, and often a need category. McKinsey research shows that well-executed personalization can generate revenue growth of around 10 to 15% (or even more depending on the sector), while failures in personalization are costly in terms of trust. On the expectations side, Salesforce research on the connected customer reminds us that a majority of consumers expect brands to understand their needs: recommendations based on purchases are a direct lever to meet that expectation. Finally, Statista surveys on personalized brand experience confirm global interest in tailored journeys, with a significant share of respondents favoring companies that treat them as individuals. Here's how to structure a recommendation strategy based on history, without sacrificing compliance or user experience.
Estimated reading time: 14 min
Summary
What is a recommendation based on purchase history?
This is a personalized suggestion based on past orders: purchased items, categories, average basket value, frequency, possible seasonality. The goal is not only to display « others also bought », but to adapt the logic to each profile: complementarity (cross-sell), replenishment (repeat purchase), upselling, or discovery guided by buyers with similar tastes.
In practice, quality takes precedence over quantity: better a few well-calibrated suggestions than ten unconvincing items. Purchase data must be cleaned (duplicates, cancellations, internal tests) so as not to distort the models. E-commerce teams that document their exclusion rules and their sources of truth (ERP, OMS, Shopify) achieve lasting gains in accuracy. This groundwork also facilitates internal audits and exchanges with the DPO when personalization is based on customer profiles.
Why purchase history often beats browsing alone
Browsing and clicks are useful, but noisy: search, hesitation, comparison. A purchase validates a choice and reduces ambiguity. Models often combine:
Collaborative filtering : « buyers similar to you liked » from purchase or basket matrices.
Content-based filtering : product attributes (material, use, compatibility) to suggest logical complements.
Hybrid models : weighting between strong signals (purchase) and weak signals (view, favorite), with noise discounting.
Forrester analyses of search and product discovery emphasize the importance of tools that unify search, ranking and recommendations: purchase history precisely feeds these personalized rankings when it is properly integrated into the catalog and merchandising.
How it works: algorithms and data
In practice, your stack should expose at least: customer ID (or logged-in email), order lines, SKU, categories, price, date, status (delivered, returned). Useful returns and cancellations improve quality: avoid recommending items that are systematically returned.
Modern engines often enrich the history with real-time behavior: current cart, last session, entry channel. The web pixels and e-commerce events make it possible to link browsing and conversion for more coherent suggestions, subject to the legal framework (see GDPR section).
The main types of recommendation blocks
“Complete your collection”: same universe, same season, same product ranges.
“Frequently bought together”: basket/order co-occurrence, useful for cross-sell.
“Restock”: consumables, cosmetics, food, office supplies.
“Upgrade”: higher-end version or more complete pack for a product already purchased.
“Because you liked X”: similarity of attributes or customer segments.
New customers and “cold start”: what to do without history?
Without any history, the engine must fall back on fallback strategies: channel-segmented best sellers, category trends, current basket, recent browsing, or a short questionnaire (size, usage). From the very first order, the quality of recommendations jumps: that is why post-purchase email and the account page are priority placements. For broader tactics, connect this logic to loyalty programs that gradually identify preferences.
Concrete examples by sector
Fashion and accessories
After a summer dress: sandals, belt, bag in the same palette. After a coat: scarf and gloves for winter. Think of season and weather as scoring variables.
Beauty and perfumery
Face cream purchased: complementary serum or routine “day / night”. Repurchase cycles are short: triggering a reminder before stock runs out increases repeat purchases.
Food and beverages
Product pairings (cheese / wine), complementary baskets, restocking on recurring items. Beware of alcohol legal constraints depending on the country.
Electronics and leisure
Compatible accessories: case, extended warranty, consumables. Upselling works if the value is clear (battery life, protection, performance).
Home and furnishings
After a sofa: cushions, throw, matching coffee table. Visualize bundles by room to make discovery easier.
For a global view of suggestion strategies, read how to increase sales with intelligent product recommendations.
Setting it up: steps and channels
Data quality first: clean categories, complete product attributes, variant management (size, color), exclusion of out-of-stock items.
Identify placements: post-purchase emails (often the best contextual ROI), account page, product page, cart, chat pop-in.
Choose the engine: native Shopify features, specialized apps, or a conversational AI layer that combines history and in-session intent.
Business rules: minimum margins, brand exclusions, respect for seasons, no duplicates with the last purchase unless a scheduled repurchase.
Measurement and iteration: CTR dashboards, assisted conversion, attributed revenue, A/B tests on titles and number of items.
Shopify’s guide on AI recommendation systems highlights the value of combining online behavior and implicit preferences to bring the experience closer to in-store advice: purchase history is a central building block of these implicit preferences.
Split between email, site, and messaging
Transactional email (confirmation, shipping) shows high open rates: it is an ideal moment for relevant add-ons, provided the message does not turn into an unreadable catalog. Automated “post-purchase” marketing campaigns make it possible to iterate on the creative and the offer, subject to communication preferences and the legal framework. On the site, the account page and product page capture active intent: this is where you test more layout variations.
Messaging (chat, messaging) adds a conversational layer: instead of a static block, the customer asks a question (“what size goes with the model I bought?”) and the system cross-references history and catalog. This is where platforms like Qstomy bring support and merchandising closer together. Statista surveys on the most valued aspects of a personalized experience show the importance of tailored offers and journeys: your channel split must reflect these expectations without overwhelming the customer.
In B2B or for complex baskets, segment by buyer role and by profit center if the data exists: history aggregated at company level can mask individual needs, but it remains useful for recommending consumables or recurring spare parts.
GDPR, consent and transparency
In the European Union, the use of data to personalize content and offers must be based on a clear legal basis (often the contract or legitimate interest, sometimes consent depending on the case) and on informing the individual. The CNIL recalls the principles of minimization, fairness, and transparency: indicate how purchases are used for personalization, offer a simple way to refuse or limit certain processing where the law requires it, and secure access. Non-essential trackers and advertising profiles are more sensitive: separate analytics, email marketing, and onsite recommendations for clear policies.
Metrics and experimentation
CTR on blocks « based on your purchases » and comparison with generic blocks.
Post-click conversion rate and attributed revenue (last click vs assisted position depending on your model).
Average order value and items per order when the recommendation is present.
Retention: impact on repeat purchases at 30 / 60 / 90 days for repurchase scenarios.
A/B testing, calibration and frequency
Recommendations based on history are not set once and for all: they require controlled experiments. In email, test the subject line, number of products (three versus six), and delay after delivery (day 2 versus day 7). On the site, compare above or below the fold, carousel versus grid, and wording of the block title. Measure revenue per open or per session, not just CTR: a more aggressive subject line may inflate clicks while degrading the quality of product visits.
Frequency is critical: too many "inspired by your purchases" messages become tiresome, especially if the catalog is shallow. For long-cycle categories (furniture, equipment), favor seasonal triggers. For consumables, pace reminders aligned with estimated usage duration. McKinsey summaries on the value (or cost) of personalization give the business order of magnitude: your job is to translate it into quantified monthly iterations.
Finally, cross results with margin: a recommended bestseller can lift CTR without improving contribution. Introduce margin floors or turnover targets where exclusion rules allow it.
Market studies on personalization cited by McKinsey also distinguish the leaders: those who excel capture a significantly higher share of revenue from targeted programs. Your dashboards should bring these macro orders of magnitude closer to your segments (new versus repeat, high basket versus low), as Salesforce customer analyses on the expectation of individualized experiences suggest.
Benefits for conversion and loyalty
Relevance : less noise, more trust.
Cross-sell and upsell : higher average order value when add-ons are useful.
Repeat purchase : automation of reminders for consumables.
Differentiation : a « advisor » experience close to premium retail.
Continuous learning : email and onsite campaigns fuel improvement loops.
Salesforce surveys highlight the importance of the overall experience : a well-calibrated recommendation supports brand perception, not just a single click.
Best practices and common mistakes
Best practices
Exclude recent one-off purchases of the same SKU to avoid the “already bought” effect.
Prioritize post-delivery emails: the customer has the product in hand, and usage momentum is strong.
Segment: a one-time customer receives different catalog entries than a VIP with twelve orders.
Test headlines and formats: “Complete your outfit” vs “Inspired by your last order”.
Common mistakes
Out-of-stock products: always synchronize inventory and lead times.
Too many identical blocks on the same page: fatigue and dilution.
Ignoring returns: risk of recommending items that often end up in customer service.
Over-segmenting without volume: statistically fragile; keep robust fallback groups.
Internationally, adapt recommendation rules to local assortments and logistics lead times: a purchase history in the German market should not trigger suggestions that are unavailable on the French site if inventories are not unified. The same principles apply to currencies and national promotions: filter products eligible for the customer’s default shipping country. Clear governance between merchandising, CRM, and the data team avoids inconsistencies visible to the customer.
Qstomy: AI, chat and purchase history
Solutions like Qstomy combine conversational assistance with product discovery logic: the chat can leverage store context (catalog, inventory) and behavior to suggest relevant add-ons, while purchase history informs suggestions when the customer is identified. The challenge is twofold: relieve support teams of repetitive questions and increase order value through contextual recommendations. For Shopify, AI chatbot integration lets you connect this experience directly to your store.
Summary
Recommendations based on purchase history rely on strong signals to go beyond simple bestsellers. Combine collaborative filtering and product attributes, handle the cold start, place blocks on high-context channels (post-purchase email, customer account), measure CTR and attributed revenue, and frame data processing within a transparent GDPR framework. McKinsey, Salesforce, Statista, and Forrester references converge: personalization is a growth lever, provided data is reliable and execution is continuous. With Qstomy, you connect support and suggestions in a seamless journey.
FAQ
Does history suffice without navigation?
Not ideally: purchase is the most reliable signal, but navigation and the cart capture immediate intent. High-performing models weight both and reduce the weight of isolated clicks.
How do you handle B2B customers or multiple buyers on the same account?
Separate profiles if possible (sub-accounts, roles) or use conservative rules: recommendations by purchase category rather than by a fuzzy individual.
Do you need a heavy CRM?
Not necessarily: the key is access to orders and the catalog. A CRM enriches segments, but the store may already be enough for a first level.
What frequency for history-based emails?
Avoid over-solicitation: favor triggers (delivery, end of product cycle) rather than a fixed weekly cadence identical for everyone.
How can you limit biases (filter bubble)?
Inject some discovery: new items, editorial picks, adjacent categories to broaden the catalog.
What legal bases for personalization in the EU?
This depends on the processing: performance of the contract, balanced legitimate interest, or consent for certain tracking or advertising profiling uses. Document and inform, as European guidelines and the CNIL remind us.
Do historical recommendations always beat general trends?
For identified customers with multiple purchases, yes on average for relevance. For cold traffic, trends and bestsellers remain essential fallback options.
Learn more
March 12, 2025
