.png)
A real-time recommendation engine lifted average order value by 31% across a catalogue of 480,000 SKUs within 6 weeks of going live.
A mid-market online retailer selling across home, garden, and lifestyle categories had built a catalogue of 480,000 active SKUs over 11 years of trading. Their product recommendation logic consisted of 14 manually maintained merchandising rules — "customers also bought" carousels populated by the e-commerce team based on intuition and seasonal judgement, updated quarterly when someone had time to review them. Average order value had been flat for 18 months at $67.40. Items per order sat at 1.8 across all categories. The merchandising team knew the recommendations were underperforming but had no data infrastructure to measure by how much or diagnose why.
Verttx built and deployed a real-time AI recommendation engine trained on 34 months of transaction history and integrated across five recommendation surfaces on the retailer's e-commerce platform. The engine went live 6 weeks after the initial discovery call. Average order value rose to $88.30 within the first month of full operation — a 31% lift.
The 14 merchandising rules had been written by two people who had since left the business. Nobody on the current team fully understood the logic behind several of them, and three rules were actively contradicting each other — recommending products in one context that were excluded from recommendations in another based on category assignments that had been reorganised since the rules were written. The e-commerce platform's built-in recommendation widget was pulling from these rules in a sequence that nobody had audited in over a year.
The data told a clear story about the cost. An analysis of 2.3 million orders from the previous 18 months showed that customers who interacted with a recommendation carousel — clicking through to a suggested product — had an average order value of $94.20, compared to $61.80 for customers who did not. The problem was not that customers would not respond to recommendations. It was that the recommendations they were seeing were irrelevant 68% of the time — a figure derived from the carousel click-through rate of 3.2% against the industry benchmark of 8-12% for well-optimised recommendation surfaces.
The catalogue size made manual curation untenable at any meaningful scale. With 480,000 SKUs, producing relevant recommendations across all product combinations by hand was not a resource problem — it was structurally impossible. The retailer needed a system that could learn from purchase behaviour at catalogue scale and surface recommendations that were genuinely relevant to the individual shopper in real time.
Verttx designed a hybrid recommendation system combining collaborative filtering, content-based filtering, and a contextual re-ranking layer. No single recommendation approach performs well across all catalogue sizes, shopping contexts, and customer lifecycle stages — particularly for a retailer with 480,000 SKUs where catalogue sparsity makes pure collaborative filtering unreliable for newer products with limited purchase history.
The collaborative filtering component identifies purchase pattern similarities across the customer base — customers who bought Product A and Product B also tend to buy Product C — using an implicit feedback matrix factorisation model trained on 34 months of transaction data covering 6.8 million orders. The content-based component uses product attribute similarity — material, colour family, style category, price tier, and brand — to recommend products that share meaningful characteristics with items the customer has viewed or purchased, handling the cold-start problem for new SKUs that have insufficient purchase history for collaborative signals. The contextual re-ranking layer adjusts the candidate set generated by both models in real time based on eight session-level signals: current cart contents, browsing path in the current session, time of day, device type, referral source, geographic region, promotional price eligibility, and current inventory status.
The engine was integrated across five distinct recommendation surfaces, each with a different primary objective:
Before full deployment, all five surfaces were A/B tested against the existing merchandising rules in a 50/50 traffic split for two weeks. The A/B test confirmed statistically significant lifts across all five surfaces at a 95% confidence level before the full cutover was approved. The test also revealed one counter-intuitive finding: on the cart page, the AI recommendations outperformed the rules for add-to-cart rate but underperformed for average added item value. A price floor constraint was added to the cart surface — preventing recommendations below 60% of the median cart value — which resolved the discrepancy without material impact on attach rate.
Within six weeks of full deployment across all five surfaces, average order value had risen from $67.40 to $88.30 — a 31% lift. Items per order increased from 1.8 to 2.6. Recommendation carousel click-through rate increased from 3.2% to 9.4% — moving from below-benchmark to above-benchmark for the retailer's category mix. The cart page attach rate — the percentage of checkouts where at least one recommended item was added — rose from 8.1% to 22.7%.
Over the first six months of full operation the engine generated $4.1 million in incremental revenue — revenue attributable specifically to recommendation-driven orders above the pre-implementation baseline. The post-purchase email surface, which had the longest feedback loop, reached full performance at month three and became the highest-ROI surface by month five: customers who engaged with the post-purchase recommendations had a 34% higher 90-day repeat purchase rate than the control group who received the previous static email.
The merchandising team, previously spending approximately 12 hours per week maintaining the 14 manual rules, now spends that time on strategic catalogue decisions — supplier negotiations, category expansion, and promotional planning — work that requires human judgement and that had been crowded out by rule maintenance. The 14 rules have been retired entirely.
The full recommendation system — all five surface integrations, the collaborative and content-based models, the contextual re-ranking layer, and the A/B testing infrastructure — was transferred to the retailer's engineering team at handover, giving them complete ownership to iterate, retrain, and extend the system independently.
We knew our recommendations were bad. We just didn't know how bad until we saw the A/B test data. The difference between what the rules were doing and what the AI was doing was not marginal — it was embarrassing. The cart surface alone paid for the engagement in the first eight weeks. — Head of E-commerce, Online Retailer
.png)
Average order value rose from $67.40 to $88.30 — a 31% lift — within six weeks of full deployment. Items per order increased from 1.8 to 2.6. Recommendation carousel click-through rate moved from 3.2% to 9.4%, above industry benchmark. Cart attach rate rose from 8.1% to 22.7%. The engine generated $4.1 million in incremental revenue in the first six months. The merchandising team retired all 14 manual rules and redirected 12 hours per week to strategic catalogue decisions.
.png)
.png)
