Ecommerce personalization: show each shopper what they will buy next.

McKinsey's 2023 research found that fast-growing companies derive 40% more of their revenue from personalization than slower-growing peers, that 71% of consumers expect personalized interactions, and that 76% get frustrated when they do not find them. Epsilon's survey put it simply: 80% of consumers are more likely to buy from brands that personalize.

On the store itself, Clerk.io reports that product recommendations draw just 7% of ecommerce traffic but produce 24% of orders and 26% of revenue. Email is still the highest-return owned channel: the DMA measures $36 back for every $1 spent, with 77% of that return coming from segmented and triggered sends, not newsletters.

And yet the average repeat purchase rate is 28.2%, according to Shopify. If most stores personalize and most customers still do not come back, most personalization is not doing what the word promises.

In 2016, Covington, Adams and Sargin at Google published "Deep Neural Networks for YouTube Recommendations". It describes a two-tower model: place shoppers and items in the same vector space, measure similarity, return the closest items. Nearly every personalization vendor built on that blueprint.

The method answers one question: what is most similar to what this person already did? It optimizes for the click. A discount buyer gets more discounts. A bargain hunter sees more bargains. The shopper who might have bought a mid-range product never sees it, because the engine has no idea what a customer is worth or what they will do next.

Research backs the gap. Theocharous, Thomas and Ghavamzadeh showed at IJCAI 2015 that a recommendation policy with a lower click rate produced more revenue and higher lifetime value than the policy that maximized clicks. The ACM RecSys 2023 tutorial on customer lifetime value states that optimizing long-term value is still hard because systems optimize clicks, ratings and dwell time. Our recommendation engine guide goes deeper.

What the agent does instead

The agent starts from one profile per shopper, including visitors who never log in. It predicts the next basket and when it will be bought, and it picks the action that moves each person toward the goal you choose. Conversions, basket size or lifetime value: you set it, it decides per person. A holdout group without the agent shows the real lift (how it decides).

1. On the site

Search that ranks results for the person typing, recommendations on every page with a different job per page (home, category, product, cart), and banners that change with the shopper's state. In a kids' fashion store, the 4.9% of visitors who used search produced 27.5% of revenue (case study).

2. In email and messages

The point is not the first name in the subject line. It is which products the email picks and when it goes. One womenswear brand moved its abandoned-browse flow to the agent and added recommendations: 3 times the shoppers reached and 9.4 times the revenue on the same flow, in a matched month (case study).

3. In segments

Recency, frequency and value say where a customer is. Predicted value says where they are going. Behavior says how to reach them. The segments worth acting on combine all three: top customers past their usual buying interval who browsed a category and did not buy.

4. In the ad audiences

What you send back to Meta and Google is personalization too. Audiences built from predicted customer value, with purchases counted once and existing customers excluded, tell the ad platform who to find (ads audiences).

5. In the loyalty program

Points and vouchers are a signal, not just a reward. Who redeems, what they redeem and who lets points expire should feed the same profile that picks products and times messages (loyalty guide).

6. In the product data

Every recommendation is only as good as the catalog behind it: stock, sizes, prices and attributes as they are now. The product database is part of the foundation in every plan.

Does it know what a customer is worth, or only what they clicked? Can it recommend something the shopper has never browsed? Do the email picks differ from the on-site picks? Does a first-time visitor get treated differently from your best customer? Can you measure the lift with a holdout, not just the click rate?

If the answer is no to three or more, the engine retrieves. It does not personalize. In a 50/50 test at a kitchen appliance store, the agent produced 2.3 times as many add-to-carts per visitor, significant at 95% (how we test).