ADA announced its acquisition of Bengaluru-based Algonomy on July 30 after the retail AI company had been used by more than 400 brands. Algonomy reports that Aditya Birla Fashion and Retail increased recommendation-driven average order value by 13%. The interesting part is not the acquisition. It is that retailers are treating product ranking as a revenue system instead of a static website feature.
That shift matters when customer acquisition keeps getting more expensive. If a store already pays to bring 100,000 people to its site each month, improving what those visitors see can be more economical than buying another 10,000 clicks.
We think many retailers are approaching this backward. They start with a chatbot because it looks like AI. The better first project is often much less theatrical: put the right products in front of each shopper, then measure whether more people buy.
What ecommerce personalization actually changes
Ecommerce personalization adjusts a store for the person visiting it. The system can reorder products, improve search results, recommend related items, or choose an offer based on signals such as browsing behavior, purchase history, location, inventory, and the device being used.
A recommendation engine is the software that makes those choices. It looks for patterns in past behavior and product data, then predicts which items are most relevant in a particular moment. This is more useful than a fixed "customers also bought" block because the answer can change by shopper and context.
Consider two people opening the same footwear category. One has repeatedly viewed running shoes in size 9. The other has bought formal shoes and filtered for leather. A personalized store should not make both people work through the same product order.
The goal is not to make the website feel clever. It is to reduce the distance between intent and a useful product.
The revenue case can be calculated before you build
Suppose an online retailer gets 100,000 monthly sessions, converts 2% of them, and has an average order value of ₹3,000. That produces ₹6 million in monthly revenue.
If personalization raises average order value by 10% while conversion stays unchanged, monthly revenue rises by ₹600,000. That is ₹7.2 million over a year before implementation costs.
This is a model, not a forecast. More revenue does not automatically mean more profit. Recommendations can shift buyers toward discounted products, increase returns, or promote an item that would have sold anyway.
That is why we would track contribution margin and return rate alongside conversion and average order value. A system that produces more orders while eroding margin is not working.
Algonomy's published case studies provide useful reference points, with the usual caveat that they come from the vendor. It reports a 13% increase in recommendation-driven average order value for Aditya Birla Fashion and Retail, and a 17% increase in average order value for a Danish retailer. Algonomy also cites McKinsey research suggesting that strong personalization can lift revenue by 5% to 15%.
Those figures are not guarantees. They are enough to justify a controlled test.
Start with one decision that affects revenue
Retailers do not need to personalize every page on day one. In fact, trying to do that usually delays the first useful result.
Pick one decision with enough traffic to measure:
-
The order of products on a category page
-
The first page of search results
-
Recommendations on a product or cart page
-
A replenishment email based on the expected life of a purchased item
Then split eligible visitors into two groups. One sees the current experience. The other sees the personalized version. This is an A/B test, a direct comparison designed to show whether the change caused the result.
Keep a small control group even after the first win. Seasonal demand, promotions, and competitor pricing can all make revenue move. Without a control group, a team can easily give its model credit for a festival sale.
Good data matters more than a fashionable model
Personalization projects are often sold as model-selection exercises. That is the wrong emphasis.
The system needs a reliable product catalog, current inventory, click and cart events, completed purchases, and returns. If product attributes are inconsistent or purchase events fire twice, a sophisticated model simply makes bad decisions faster.
Inventory is especially important. Recommending an unavailable size or a product that cannot reach the buyer's location wastes attention and damages trust. For retailers with thin margins, the ranking should also understand margin, stock age, and fulfillment cost rather than optimizing for clicks alone.
New visitors create another limitation. The system has no personal history for them. This is called the cold-start problem. A practical fallback uses context, popular products, and product attributes until the visitor provides enough signals.
Personalization should also respect consent. A business needs clear rules for what data it collects, how long it keeps that data, and which decisions use it. More tracking is not always better. If the experience feels invasive, the retailer has traded a possible conversion gain for a trust problem.
Buy the commodity layer, build the advantage
A retailer on a standard Shopify setup with a clean catalog may be better served by an established personalization app. Paying a monthly fee is usually more sensible than funding a custom recommendation platform.
Custom development starts to make sense when the business has unusual constraints. Examples include region-specific inventory, complex bundles, marketplace sellers, offline purchase data, or pricing rules that a packaged tool cannot represent. It can also be justified when recommendation logic is central to how the company competes.
Even then, building everything from scratch is rarely wise. We prefer to use proven infrastructure for event collection and model serving, then build the business-specific ranking rules, integrations, and measurement layer around it.
The first version should be narrow enough to ship in weeks, not quarters. If it cannot beat the existing experience in a fair test, stop or change direction. There is no prize for keeping an AI project alive after the economics fail.
The acquisition is a signal, not a strategy
ADA says the combined business will operate across 34 markets. That scale shows where retail technology is heading: decisions about products, offers, and messages are becoming more responsive to each customer's context.
But copying a large retailer's stack is not a strategy for a growing business. The sensible move is to identify one high-volume customer decision, connect it to reliable data, and prove the financial result.
That is the work we do at Axentia. We build AI systems and full-stack products around the real constraints of a business, including catalog quality, integrations, measurement, and operating cost. If you want to test whether ecommerce personalization can improve your store's economics, book a call with us.