AI Engineer World's Fair 2026
Keyword Search Is Dying. Is Your Catalog Ready for AI Agents? — PayPal
Loading slides…
Read the talk
Keyword Search Is Dying. Is Your Catalog Ready for AI Agents?
Nixon Dinh explains how shopping moves from keywords to intent and delegated action, then follows PayPal’s catalog-enrichment experiments to a practical finding: useful product detail improves recommendations, while excess text can make them worse.
From a talk by Nixon Dinh
At a glance
Ideas worth remembering
Intent-driven shopping makes product context more important: agents must connect a shopper’s situation to suitable items, rather than depend only on matching keywords.
Hybrid search combines predictable literal matching with broader meaning-based retrieval; the recording does not disclose PayPal’s ranking implementation.
PayPal’s experiments found the largest enrichment gains in thin catalogs. Product attributes, deeper descriptions, buyer context and reviews can supply information those records lack.
Over-enrichment sometimes increased hallucinations and reduced performance. Assess the starting catalog and add relevant product information rather than boilerplate.
From finding a product to describing a need
A shopper typing “school supplies” has already translated a situation into a search term. A parent explaining that her daughter starts fifth grade in September gives an agent the situation itself. The system now has to connect that need to products. Nixon Dinh, Director of Product for Agentic Commerce at PayPal, uses this change to introduce the catalog problem: merchants need product data that helps an agent understand what an item is useful for.
Commerce has already expanded from physical stores to e-commerce sites, then into email, SMS and social channels. PayPal sees agents as the next step in that progression. The change reaches beyond where a merchant advertises: it changes how the customer searches and how products become candidates for a purchase.
The progression has three stages:
- Search era. The shopper enters keywords and navigates a catalog. Merchants optimize for SEO, while the shopper carries the work of finding a suitable product.
- Intent era. The shopper describes a need in ordinary language. AI must connect that description to relevant products, bringing brands into the interaction.
- Delegation era. As customers come to trust recommendations, PayPal expects them to hand agents more shopping and commerce tasks. This stage is a forecast, and trust is the condition that makes it plausible.
Discoverability changes before delegation arrives
Dinh cites a Bain & Co. projection that agents will execute 15–25% of all commerce by 2030, alongside a prediction that agents will outnumber humans online within a decade. He also presents expectations of $1 trillion in agentic US retail commerce, $3–5 trillion globally, and more than two billion people starting shopping journeys with AI over the next few years. These are forecasts used to frame the opportunity, rather than measurements of commerce already delegated to agents.
The nearer-term signals concern shopping assistance. PayPal’s consumer studies found that 39% of US shoppers already used AI for shopping, and Dinh reports that businesses saw up to four times higher order-completion conversion when an interaction began with an agent. He connects that result to richer product information available during the interaction. The recording does not establish the study populations or comparison conditions, so the conversion figure is a reported observation rather than a general causal estimate.
For a merchant, each stage introduces a different job. Keyword SEO might make a summer dress discoverable in a search. Intent-driven shopping requires the agent to recognize when that dress fits a customer’s request and represent it usefully. Delegation adds the requirement to turn demand into a transaction safely and in a trusted way. Catalog preparation matters before that final stage: a product cannot become a good recommendation if the agent cannot discover or understand it.
An advertising feed does not explain product fit
Many merchants already syndicate catalogs to Google, Meta or other advertising platforms. That does not make those catalogs ready for agents. Their specifications were built around existing advertising and human-search workflows, and may provide too little readable context for an agent to discover products or judge their relevance to a request.
There is also no single recommendation system to optimize against. Agents can use different models, ranking methods and recommendation algorithms; Dinh also notes that vector retrieval is not always deterministic. That makes a universal agent-discovery strategy harder to define than a familiar SEO strategy. PayPal’s response was to enrich merchant product data and run experiments on how agents searched and recommended products.
The distinction between keyword and semantic search explains what enrichment can change:
- Keyword matching. In the talk’s simplified example, a search for blue running shoes finds products containing those words. Literal matching is fast, precise and predictable, but “footwear for jogging” may miss the same shoes because the wording differs.
- Semantic matching. The system converts a query into vectors that represent meaning. A request for something comfortable for marathon training can lead to running shoes even without matching the query’s words literally. This broadens discovery, but introduces fuzzier relationships between the request and the product.
PayPal’s direction is hybrid search: combine literal matching with meaning-based retrieval. The two methods supply different strengths, so improving semantic discovery does not eliminate the usefulness of keywords. Dinh does not specify how PayPal combines scores or ranks the combined results; the claim here is about the complementary approaches, rather than a disclosed ranking implementation.
Give the running shoe enough detail to match a need
Enrichment starts with the catalog a merchant actually has. A thin catalog and a specification-heavy catalog need different additions. PayPal considered filling missing attributes, deepening descriptions, adding buyer context, including trust signals and clarifying product identity. These are different ways to make a product understandable; adding the same material to every catalog would ignore what is already present.
The running-shoe example makes the change concrete. The original record identifies a blue men’s shoe, model 8543. Its description says blue running shoes are available in multiple sizes, and its attribute identifies the color as blue. That record gives keyword search a few useful terms—blue, running and shoe—but gives semantic retrieval little context about comfort, construction or intended use.
The enriched version describes a men’s lightweight running shoe. The description adds cushioning and breathability; attributes add mesh material and a true-to-size fit; review data supplies another source of product context. Search can now match a broader set of terms, including running, training and mesh. Meaning-based retrieval also has more product characteristics to connect with a shopper’s request.
Why can the same shoe become easier to find after enrichment? The diagram follows the change in its representation. The product has not changed; the catalog has made more of its characteristics available to search. Cushioning and breathability give a comfort-oriented request something relevant to match, while explicit attributes widen literal matches. The example illustrates that connection without reporting a recommendation score for this individual shoe.
The same shoe, with more searchable product information
Read the diagram as text
- Thin shoe record. Blue men’s shoe, model 8543; running shoes in multiple sizes; color: blue.
- Enriched shoe record. Lightweight, cushioned, breathable; mesh material; true-to-size fit; review data.
- Broader keyword matches. retrieval — Search can match running, training and mesh.
- Richer semantic context. retrieval — Comfort and construction details give meaning-based search more context for matching a need.
- Thin shoe record → Enriched shoe record: Add attributes, description detail and reviews.
- Enriched shoe record → Broader keyword matches: Expose more matching terms.
- Enriched shoe record → Richer semantic context: Supply more relevant meaning.
Enrichment helped; excess context weakened the signal
In PayPal’s head-to-head experiments, enriched product data improved recommendations compared with the unenriched baseline. Merchants with the thinnest catalogs gained the most, and weak product data left the most room for improvement. This was one set of experiments, with no quantified recommendation metric or evaluation protocol provided in the recording; it supports trying enrichment against a merchant’s own baseline rather than assuming a universal gain.
The result depended on what the added content contributed. Product-specific detail gave search more relevant information. Unstructured boilerplate could dilute that signal. A welcome message or a statement that a store is family-owned may describe the merchant, but it does little to explain whether a particular shoe suits a training request. Putting those words into product context adds text without necessarily adding useful evidence of product fit.
In some over-enriched cases, agents hallucinated more and performed worse. That qualifies the positive result: useful enrichment beat sparse records, but continuing to add content was not a reliable improvement strategy. Keyword search benefited from added terms, while semantic retrieval required attention to how meaning was represented and how much irrelevant context entered the record.
Start with missing product information
The practical starting point is to assess the catalog, then choose an enrichment strategy for its weaknesses. A sparse shoe record may need material, fit and comfort details. A specification-heavy catalog may already have those facts and need a different treatment. The experiments offer a direction—structured, high-quality product content—without a concrete prescription for every business.
“More text isn’t better” is the useful closing rule. Dinh’s view is that content quality will matter more than brand identity in agent discovery, while other trust signals still influence recommendations. Reviews in the running-shoe example show how those concerns can meet: they add context an agent can use alongside the product’s attributes. The merchant’s job is to give an agent enough relevant information to make a recommendation worth trusting.
Full transcript
Loading transcript…