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 sources for personalizing recommendations: a purchase confirms an intention, a budget, and often a category of need. McKinsey's work shows that well-executed personalization can generate revenue growth on the order of 10 to 15% (or even more depending on the industry), while failures in personalization are costly in terms of trust. In terms of expectations, 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 for meeting 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 is 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 recommendation based on past orders: purchased items, categories, average basket value, frequency, and possible seasonality. The goal is not just to display “others also bought,” but to adapt the logic to each profile: complementarity (cross-sell), replenishment (repeat purchase), premium upgrade (upsell), or discovery guided by buyers with similar tastes.
In practice, quality matters more than quantity: a few well-calibrated suggestions are better than ten barely credible items. Purchase data must be cleaned (duplicates, cancellations, internal tests) so as not to skew the models. E-commerce teams that document their exclusion rules and their sources of truth (ERP, OMS, Shopify) achieve lasting gains in accuracy. This foundational work also makes internal audits easier and facilitates discussions with the DPO when personalization relies on customer profiles.
Why purchase history often beats browsing alone
Browsing and clicks are useful, but noisy: searching, hesitation, comparison. A purchase validates a choice and reduces ambiguity. Models often combine:
Collaborative filtering: “buyers similar to you also liked” based on 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 on product search and discovery emphasize the importance of tools that unify search, ranking, and recommendations: purchase history feeds these personalized rankings precisely 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 identifier (or logged-in email), order lines, SKUs, categories, price, date, status (delivered, returned). Useful returns and cancellations refine quality: avoid recommending items that are systematically returned.
Modern engines often enrich history with real-time behavior: current cart, last session, entry channel. Web pixels and e-commerce events make it possible to link browsing and conversion for more consistent suggestions, subject to the legal framework (see GDPR section).
The main types of recommendation blocks
"Complete your collection": same universe, same season, same ranges.
"Frequently bought together": cart/order co-occurrence, useful for cross-sell.
"Restock": consumables, cosmetics, food, office supplies.
"Move upmarket": higher-end version or more complete bundle for a product already purchased.
"Because you liked X": proximity of attributes or customer segment.
New customers and “cold start”: what to do without historical data?
Without history, the engine must switch to fallback strategies: channel-segmented best-sellers, category trends, current cart, recent browsing, or a lightweight questionnaire (size, use). From the very first order, recommendation quality jumps: this is why the post-purchase email and 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 with a winter approach. Think about season and weather as scoring variables.
Beauty and perfumery
Face cream purchased: complementary serum or a "day / night" routine. Repurchase cycles are short: triggering a reminder before depletion increases recurrence.
Food and beverages
Product pairings (cheese / wine), complementary baskets, restocking recurring references. Be mindful of legal alcohol 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 furnishing
After a sofa: cushions, throw, matching coffee table. Visualize bundles by room to make discovery easier.
For an overall view of suggestion strategies, read how to increase sales with smart product recommendations.
Implementation: steps and channels
Data quality first: clean categories, complete product attributes, variant management (size, color), out-of-stock exclusions.
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 cross-references history and in-session intent.
Business rules: minimum margins, brand exclusions, season compliance, no duplicates with the latest purchase except for scheduled repurchase.
Measurement and iteration: CTR dashboards, assisted conversion, attributed revenue, A/B tests on titles and number of items.
Shopify's guide to 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 core building block of these implicit preferences.
Distribute across email, site, and messaging
Transactional email (confirmation, shipping) shows high open rates: it is an ideal moment for relevant add-ons, provided the message is not turned into an unreadable catalog. Automated "post-purchase" marketing campaigns make it possible to iterate on creative and offer, subject to communication preferences and legal requirements. On-site, the account page and product page capture active intent: more layout variants can be tested there.
Messaging (chat, messaging) adds a conversational layer: instead of a static block, the customer asks a question ("what size 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 should reflect these expectations without overwhelming the customer.
In B2B or for complex carts, segment by buyer role and profit center if the data exists: history aggregated at the company level may hide individual needs, but it remains useful for recommending recurring consumables or 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 reiterates the principles of data minimization, fairness, and transparency: indicate how purchases are used for personalization, offer a simple way to refuse or limit certain processing when required by law, and secure access. Non-essential trackers and advertising profiles are more sensitive: separate analytics, email marketing, and onsite recommendations for readable policies.
Metrics and experimentation
CTR on “based on your purchases” blocks and comparison with generic blocks.
Post-click conversion rate and attributed revenue (last click vs assisted position according to your model).
Average basket and items per order when the recommendation is present.
Retention: impact on repeat frequency at 30 / 60 / 90 days for repurchase scenarios.
A/B testing, calibration and frequency
Recommendations based on historical data are not set-and-forget: they require controlled experiments. In email, test the subject line, the number of products (three versus six), and the delay after delivery (D+2 versus D+7). On-site, compare above versus below the fold, carousel versus grid, and the wording of the block title. Measure revenue per open or per session, not just CTR: a more aggressive subject line can inflate clicks while degrading the quality of product visits.
Frequency is critical: too many “inspired by your purchases” messages cause fatigue, especially when the catalog lacks depth. For long-cycle categories (furniture, equipment), favor seasonal triggers. For consumables, schedule reminders aligned with the estimated usage duration. McKinsey summaries on the value (or cost) of personalization provide the business order of magnitude: your job is to translate it into quantified monthly iterations.
Finally, cross-reference results with margin: a recommended best-seller can drive CTR without improving contribution. Introduce margin floors or turnover objectives when exclusion rules allow it.
The market studies on personalization cited by McKinsey also distinguish leaders: those who excel capture a significantly higher share of revenue from targeted programs. Your dashboards should connect these macro orders of magnitude to your segments (new versus returning, high versus low basket), as suggested by Salesforce customer analyses on the expectation of individualized experiences.
Benefits for conversion and loyalty
Relevance: less noise, more confidence.
Cross-sell and upsell: a higher average basket when complementary products are useful.
Repeat purchase: automated reminders for consumables.
Differentiation: a “advisor” experience close to premium retail.
Continuous learning: email and onsite campaigns feed improvement loops.
Salesforce surveys highlight the importance of the overall experience: a well-calibrated recommendation supports brand perception, not just an isolated click.
Best practices and common mistakes
Best practices
Exclude recent non-recurring purchases of the same SKU to avoid the “already purchased” effect.
Prioritize post-delivery emails: the customer has the product in hand, and usage momentum is strong.
Segment: a customer with one purchase receives different catalog entries than a VIP with twelve orders.
Test titles and formats: “Complete your outfit” vs. “Inspired by your last order”.
Common mistakes
Out-of-stock products: always keep inventory and lead times synchronized.
Too many identical blocks on the same page: fatigue and dilution.
Ignoring returns: risk of recommending items that often end up in after-sales service.
Over-segmenting without volume: statistically fragile; keep robust fallbacks.
For international operations, adapt recommendation rules to local assortments and logistics lead times: a purchase history on the German market should not trigger unavailable suggestions on the French site if inventory is not unified. The same principles apply to currencies and national promotions: filter products eligible for the customer's default delivery country. Clear governance between merchandising, CRM, and the data team prevents inconsistencies visible to the customer.
Qstomy: AI, chat and purchase history
Solutions like Qstomy combine conversational assistance with product discovery logic: the chat can rely on 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 from repetitive questions and increase order value through contextualized recommendations. For Shopify, AI chatbot integration makes it possible to connect this experience directly to your store.
Summary
Recommendations based on purchase history rely on strong signals to go beyond simple best-sellers. Combine collaborative filtering and product attributes, manage 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. References from McKinsey, Salesforce, Statista, and Forrester converge: personalization is a growth driver, provided data is reliable and execution is continuous. With Qstomy, you connect support and suggestions in a smooth journey.
FAQ
Is history enough without browsing?
Ideally no: purchase is the most reliable signal, but browsing and cart capture immediate intent. High-performing models weight both and reduce the impact of isolated clicks.
How should B2B customers or multiple buyers on the same account be handled?
Separate profiles where possible (child accounts, roles) or use conservative rules: recommendations by purchase category rather than by unclear individual.
Do you need a heavy CRM?
Not necessarily: what matters most is access to orders and the catalog. A CRM enriches segments, but the store alone may already be enough for a first level.
How often should history-based emails be sent?
Avoid over-solicitation: prioritize triggers (delivery, end of product cycle) rather than a fixed identical weekly cadence for everyone.
How can bias (filter bubble) be limited?
Inject a share of discovery: new arrivals, editorial selections, adjacent categories to open up the catalog.
What legal bases apply to personalization in the EU?
It depends on the processing: performance of a contract, balanced legitimate interest, or consent for certain tracking or advertising profiling uses. Document and inform, as reminded by European guidelines and the CNIL.
Do history-based recommendations always outperform general trends?
For identified customers with multiple purchases, yes on average in terms of relevance. For cold traffic, trends and best-sellers remain essential fallbacks.
Go further
March 12, 2025
