Shopping is moving from keyword search, where shoppers find products themselves, toward intent-based interactions where they describe their needs to an AI agent.
2
Merchant catalogs syndicated to Google, Meta, and advertising platforms were built for human search and often lack the structure and context agents need.
3
PayPal's experiments found that enrichment improved recommendations, especially for thin catalogs, while excessive unstructured content diluted the signal and increased hallucinations.
Summary
Nixon Dinh describes three stages of commerce: the search era, the intent era, and the delegation era. Keyword search makes shoppers find products through literal terms, while intent-based shopping lets them describe a situation or need in their own words. As shoppers trust agents, they may delegate more shopping tasks to them. This shift creates a catalog problem. Product feeds built for advertising and SEO are readable to people and traditional search systems, but they were not designed for agent discovery and recommendation. Dinh compares keyword and semantic search, then argues for a hybrid approach. PayPal's experiments enriched product data with attributes, descriptions, buyer context, product identity, and trust signals such as reviews. Enrichment improved recommendations, with the largest gains coming from thin or weak catalogs. Adding irrelevant boilerplate had the opposite effect. The practical advice is to assess the catalog first, then add structured, useful context without assuming that more text will produce better results.
Commerce is moving from search to intent and then delegation
Dinh describes commerce as moving through three stages. In the search era, shoppers type keywords, navigate catalogs, and rely on SEO. In the intent era, they describe their situation, such as needing back-to-school supplies for a daughter starting fifth grade, and an AI system interprets the request. The delegation era begins when shoppers trust agent recommendations enough to let agents take more actions on their behalf. PayPal is building toward that delegated model by working on trust in agent-driven shopping and payments.
Agentic shopping is becoming a current merchant concern
Dinh says the delegation era may sound distant, but he presents projections that make catalog preparation relevant now. He cites a Bain & Co. study projecting that agents will execute 15 to 25 percent of commerce by 2030, with agents predicted to outnumber humans online within a decade. He also gives PayPal's estimates of $1.1 trillion in US retail commerce and $3 trillion to $5 trillion globally becoming agentic, alongside more than two billion people expected to use AI to start shopping journeys over the next few years.
Merchant catalogs were built for advertising and human search
Many merchants syndicate catalog data to Google, Meta, Facebook, or other advertising platforms. Dinh says those specifications were not built for agents. They do not necessarily make products easy for agents to discover or interpret, because they were designed around human search. This changes what merchants need to do. SEO-era work focused on ranking for terms such as a particular dress, while agentic commerce requires products to be discoverable when an agent represents the shopper's broader need.
Unknown agent models make experimentation necessary
Dinh describes the models behind agents as a black box. Agents may use different models, ranking systems, recommendation algorithms, and vector database retrieval methods, and retrieval is not always deterministic. That makes it difficult to choose one optimization strategy comparable to traditional SEO. His response is to run experiments. PayPal enriched merchant product data and tested how agents searched for and recommended the resulting products instead of assuming that one catalog formula would work everywhere.
Keyword and semantic search have different failure modes
Keyword search matches literal terms. A search for blue running shoes may return only products containing those words, so it is precise, fast, and predictable, but it can miss a request such as footwear for jogging. Semantic search converts a query into vectors based on meaning, allowing a request for something comfortable for marathon training to find running shoes even without those exact words. Dinh says semantic search can lose focus and hallucinate when too much meaning is injected. PayPal's answer is a hybrid of keyword and semantic search.
The right enrichment depends on the starting catalog
Dinh says enrichment is not one-size-fits-all. A business should first understand whether its catalog is thin or already heavy on product specifications. PayPal tested patterns including filling attributes, adding description depth, supplying buyer context, adding trust signals, and clarifying product identity. The useful additions depend on what the existing catalog lacks. A merchant with sparse product data has a different enrichment problem from one that already has detailed specifications.
A running shoe becomes more searchable through specific context
Dinh's example starts with a blue men's shoe model, a short description, and an attribute for the color. The enriched version identifies it as a lightweight running shoe and adds details such as cushioning, breathability, true-to-size fit, mesh material, and review data. The richer product can match concepts such as running, training, and mesh. The example shows how attributes, buyer context, and reviews give semantic retrieval more useful meaning than a short title and a single color keyword.
PayPal found that enrichment improved recommendations compared with an unenriched baseline, and merchants with the thinnest or weakest product data gained the most. Dinh says content quality mattered more than simply adding a schema or piling on unstructured text. Boilerplate such as welcome messages or claims that a store is family-owned may be irrelevant to an agent and can dilute the signal. In some cases, over-enrichment caused more hallucinations and poorer performance. Merchants should assess their catalog and add structured, relevant content rather than stuffing it with words.
