How Walmart Finds Lookalike Customers for E-Commerce Campaigns

A Walmart Global Tech paper presents a system for expanding marketing audiences in the United States, not a method for winning a first customer.

By Career Ahead AI · United States — AI-assisted, fact-verified against cited sources. Readers should verify key facts.

Treat audience expansion as a later-stage tool

Separate finding your first customer from finding more people like customers you already know. The research describes a system designed to expand the targeted audience of marketing campaigns, and says it can handle hundreds of millions of customers. That is evidence for a large-scale audience-expansion use: identifying people similar to customers of interest. It does not lay out how to make an initial sale, choose a first market or persuade someone to buy. If you are looking for a first-customer playbook, this system does not supply those steps; its stated purpose is broader campaign reach.^¹

The business setting is a company with a substantial customer base and existing campaigns. Customer-centric campaigns generate a large portion of Walmart’s e-commerce website traffic, while Walmart has hundreds of millions of active customers in the US market alone. Those facts help explain why audience expansion is relevant to Walmart. They do not establish that the same approach is suitable for a new business with no customer information, or that it produces a first sale. Use the research to understand a campaign-expansion method, rather than treating Walmart’s scale as evidence that every founder should adopt it.^¹

That distinction matters: a person similar on purchase activity need not be similar on visits or engagement, because the model can represent different business interests.

Decide which customer signals matter to the campaign

The system draws on several types of customer information: transactions, visits, engagements and customer metadata. This matters when deciding whether the approach fits a particular marketing objective, because it does not rely on only one kind of signal. The research describes those inputs as multimodal features from different data sources. It does not specify a minimum amount of history, a required mix of signals or a setup process for a small business. You can take away the kinds of information the model uses, but not a checklist for collecting customer data before launching your first campaign.^¹

You may also like

Entrepreneurship & Business

IIT Mandi Survey Finds 22% of Indian Students Have Started Businesses

The survey also found that 43% of students wanted to become entrepreneurs, while 35.5% of student businesses generated regular sales.

Read More →

Choose the intended business interest before interpreting a match. The paper says its similarity metric can measure interests such as purchases, visits and engagements, and describes those as meaningful goals for marketing campaigns. That distinction matters: a person similar on purchase activity need not be similar on visits or engagement, because the model can represent different business interests. The research does not rank those goals or recommend one for a particular product. It supports matching the campaign’s intended interest to a relevant signal, while leaving the choice of goal to the business.^¹

Represent customers with the signals the model supports

The proposed system represents customers with a deep-learning-based embedding model, then uses those representations to search for lookalikes. An embedding is the model’s representation of a customer as a vector, which allows the system to compare customer profiles mathematically. This is the mechanism behind its matching approach; the paper does not describe a manual list of traits for a founder to select. The distinction is useful when assessing fit: the research concerns a model-driven way to compare customer records, not a general marketing technique that works without customer data or a defined audience.^¹

The model is designed to handle both dense numerical features and sparse categorical features. The paper also notes that a customer could have hundreds or even thousands of features, which indicates the breadth of information its system is built to process. That is a technical capability, not a recommendation to gather hundreds of attributes for a new business. The research gives no required feature count, data-collection plan or cost for preparing customer records. For a founder deciding whether to apply the idea, the supported conclusion is limited: the described model handles varied feature types at high dimensionality.^¹

Use similarity search to compare customer profiles

The system estimates similarity using cosine distance between customer embedding vectors. In practical terms, it compares the direction of the model’s customer representations to estimate how alike two customers are. That provides a defined matching mechanism behind the phrase “lookalike customers”; it is not simply a claim that two people share one obvious trait. The paper does not provide a threshold for deciding when two profiles count as similar, or say how a business should choose one. A founder should therefore treat cosine distance as the method described, not as a ready-made campaign rule.^¹

For finding similar customers quickly, the research describes approximate nearest neighbor search based on FAISS. Approximate nearest neighbor search is a way to locate close matches without comparing every possible pair in a direct, exhaustive search; the paper presents it as a scalable and efficient search method. This explains how the system addresses the problem of finding matches across a very large audience. The source does not provide a price, implementation timeline or small-business setup guide, so it cannot establish the cost or effort required to use the same technology outside the described system.^¹

You may also like

Entrepreneurship & Business

TVS Motor Approves Consumer Finance Demerger

The proposed split creates two separately listed entities, with a finance-company share allocated for each TVS Motor share held.

Read More →

The paper does not provide a threshold for deciding when two profiles count as similar, or say how a business should choose one.

Keep the location-specific technique in its proper scope

The model encodes location embedding using transfer learning. That is a stated part of its customer representation, and it signals that location is included in the model’s approach rather than being described as an afterthought. The paper does not explain which locations the technique covers beyond its US-market context, or give a procedure for adapting it to another country. Do not read the location feature as proof that the model has been validated for every market. The evidence supports noting that location is encoded; it does not establish a geographic expansion plan for a small business.^¹

The research is titled “Finding Lookalike Customers for E-Commerce Marketing” and its authors are Yang Peng, Changzheng Liu and Wei Shen, all affiliated with Walmart Global Tech. Its subject is customer matching for e-commerce marketing, consistent with the system’s stated role in expanding campaign audiences. Across the described method, the evidence covers customer representations, business signals and scalable search. It does not give steps for identifying a first prospect, making an initial offer or converting a first sale. For that decision, the responsible recommendation is to use this research as a reference on audience expansion only, not to infer an acquisition plan it does not provide.^¹