Semantic Search for AI Agents: Catalogs

Semantic search for AI agents changes what a product catalog has to contain. Nixon Dinh, PayPal's director of product for agentic commerce, said in an October 2026 AI Engineer talk that enrichment improved agent recommendations in PayPal's own experiments, while unstructured boilerplate made agents hallucinate more. Most merchants still syndicate feeds written for advertising platforms, not for machines that read meaning.

Semantic search for AI agents and why catalogs need to change

Semantic search for AI agents means retrieval systems match a query against the meaning of a product listing rather than its literal words. PayPal's experiments tested whether enriching catalog data changed how agents found and recommended products, and Dinh argues the catalogs most merchants ship were built for human keyword search, which makes them harder for agents to read and rank.

Dinh frames the change as three phases of shopping behavior:

• The search era runs on keywords, catalog navigation and SEO, and businesses optimized for it for roughly a decade. It put the burden of finding products on the shopper.
• The intent era replaces typed keywords with full descriptions of a need, phrased in the shopper's own words. Instead of "school supplies," a shopper says their daughter starts fifth grade in September and they need back-to-school items.
• The delegation era follows once shoppers trust agent recommendations enough to hand over purchasing actions, which is the direction PayPal says it is building toward.

Each era asks a different question of the merchant. In the search era it was optimising for a keyword. In the intent era it is being discoverable wherever an agent acts on the merchant's behalf. In the delegation era it is converting that demand safely and in a way the shopper trusts.

The practical problem is a data problem. In the search era a merchant could optimize for a specific keyword phrase. In the delegation era, visibility depends on being discoverable by a model whose ranking logic is opaque. Dinh described the model layer as largely a black box: different agents use different models, different ranking algorithms, and vector retrieval that is not always deterministic. That removes the single clear target SEO once provided.

The recommendation from PayPal is to run experiments rather than wait for a standard. Dinh's team enriched merchant product data and measured how agents searched and recommended products, on the reasoning that an unclear optimization target still responds to testing. Dinh's framing throughout the talk was that the shift carries risk and opportunity at the same time, and that commerce has historically evolved alongside every change in consumer behavior.

Keyword search vs semantic search vs hybrid retrieval

Keyword search matches literal words and stays precise, fast and predictable; semantic search matches meaning and broadens recall, but Dinh warns that injecting too much meaning loses focus and invites hallucination. PayPal's stated direction is hybrid retrieval, which combines both signals instead of choosing between them. Dinh's one-line summary of the trade-off: keyword search is precise and literal, semantic search is smart but fuzzy.

In the transcript's own example, a query for "blue running shoes" returns only listings containing those words under keyword matching. The same catalog may fail a query for "footwear for jogging" even when the identical shoe is in stock, purely because the words differ. Semantic retrieval converts the query into vectors that represent intent, so a request for "something comfortable for marathon training" can surface a shoe that never uses those words.

The trade-off runs in the other direction too. Enrichment is the lever that moves a catalog along that spectrum, and how much to add is itself a decision point rather than a default. PayPal's hybrid approach keeps literal matching for exact terms such as a model number alongside vector matching for sentence-length requests.

How PayPal's catalog enrichment experiment worked

PayPal enriched merchant product data across several patterns and then tested whether agents recommended the enriched listings more often than an unenriched baseline. Dinh's summary of the result was direct: enrichment helped, and the unenriched baseline completely underperformed head-to-head. Dinh ran the comparison as an internal experiment across merchant catalogs, not as a controlled study on neutral data.

The enrichment patterns Dinh named cover four areas:

1. Filling attributes that were missing from the listing.
2. Adding depth to the product description.
3. Adding buyer context and trust signals, including review data.
4. Adding product identity data.

His example starts with a listing called "men's shoe model 8543" in blue with the description "I have blue running shoes" and availability in multiple sizes. Keyword matching can still hit blue, running or shoe, but semantic retrieval has almost nothing to work with.

The enriched version carries a title such as "men's lightweight running shoe" plus cushioning and breathability in the description, attributes for fit and mesh material, and review data as a trust signal. The enriched listing also added attributes such as true to size and mesh as the material. Now the retrieval layer can match running, training and mesh, among other terms. Trust signals matter to agents, Dinh said, and the mechanisms include reviews rather than brand name alone.

The boundaries of that test matter. PayPal reported this as one set of experiments within its own agentic commerce work, not an independent benchmark, and Dinh gave no universal configuration that every merchant should copy. There is also no single correct way to enrich product data for semantic search, because the right approach depends on the catalog a merchant starts from. The recommendation he did give was to assess an existing catalog first, then choose the enrichment strategy that fits it.

What the enrichment results showed across catalog types

Thin catalogs improved the most. The statement below is Dinh's own summary of running catalog enrichment experiments inside PayPal, and the section it sits in is a report of what he said rather than an independently reproduced result.

Thinner catalogs gained the most from enrichment, and weak product data had the most room to improve, while keyword search also benefits from better structured attributes.

The finding points to where effort pays off. A merchant whose listings already carry rich spec data, dimensions, materials and fit notes has less to add than one whose listings amount to a model number and a color. Dinh describes the difference between a thin catalog and a spec-heavy catalog as the starting question a merchant should answer before choosing what to enrich.

There is a second layer in the result that outlasts any single pattern. Dinh said content quality drives the outcome, not the schema, which means the same enrichment fields help or hurt depending on what fills them. Enrichment also helps keyword search, not only semantic retrieval, so a merchant who improves structured attributes sees gains even before agents become the main channel. Dinh's closing line on the research was that AI rewards content quality rather than brand, though it still optimizes for other trust signals.

Why over-enrichment makes agents hallucinate

Piling unstructured text into a product listing can lower recommendation quality. Dinh reported that in some cases over-enrichment caused the agent to hallucinate more and to underperform, which makes restraint part of the enrichment method rather than an afterthought.

The unstructured data he names is the kind a human shopper might scroll past: "welcome to our store" and "we're family-owned business" style messaging, repeated across every listing. Those lines add context for a person but no retrieval signal for an agent, and they compete with the attributes that do.

The failure mode is a side effect of how semantic retrieval works. If meaning is injected too broadly, the vector for a listing drifts away from the specific product and toward generic marketing language, and piling on unstructured data dilutes the signal. Dinh's framing puts the two failure modes side by side: too little enrichment leaves semantic retrieval with nothing to match, and too much unstructured enrichment dilutes the signal.

The conclusion he drew is blunt. More text is not better, and structured quality content is the element that wins.

PayPal's figures on agent commerce adoption

PayPal's stated projections for agent commerce remain projections, not measurements. The company expects roughly 1.1 trillion US dollars of US retail commerce to be agentic, with a global figure of 3 to 5 trillion, and cites more than 2 billion users projected to begin shopping journeys with AI over the next few years. Those numbers describe PayPal's forecast, not a recorded outcome.

Dinh also cited two figures from outside PayPal. He attributed a projection to BCG that 15 to 25 percent of all commerce would be executed by an agent by 2030, along with a prediction that agents outnumber humans online within a decade. He described 39 percent of US shoppers as already using AI for shopping today, based on PayPal consumer studies. Dinh's read of these numbers is that the shift is structural for the whole e-commerce ecosystem, not a marginal change merchants can ignore.

The conversion figure is the one most likely to be misread. Dinh said businesses interacting with an agent first tend to see up to four times higher conversion on order completion, which he linked to the richer, more in-depth product data available in that interaction. He also reported that merchants and partners described growth of over 693 percent in referral traffic from AI engines, a figure that arrives through partner reporting rather than PayPal measurement.

The gap between a projection and a result is where catalog decisions actually live, and Dinh's own advice follows that logic: assess, test, then measure against a baseline.

Trade-offs between keyword, semantic and hybrid retrieval

The table below compares the retrieval approaches on the dimensions Dinh covered in the talk. Every row reflects his description of how each approach behaves, not an independent benchmark of retrieval accuracy.

Approach	Matching basis	Main strength	Main weakness
Keyword search	Literal word overlap	Precise, fast, predictable	Misses intent when wording differs
Semantic search	Vector meaning and intent	Finds products described in new words	Loses focus and hallucinates when overstuffed
Hybrid retrieval	Keyword plus semantic signals	Keeps precision while broadening recall	Needs a catalog assessed and enriched first

The hybrid row is where PayPal says it is headed. The reason is structural rather than stylistic: keyword matching protects exact matches such as a model number or a specific color, while semantic matching handles the sentence-length requests that define the intent era.

Dinh's caution applies across all three rows. Enrichment changes how each approach performs, and the amount of meaning added is a variable a merchant controls.

How to make a catalog agent-ready

Agent-ready catalogs come from assessing what data exists today, enriching attributes, description depth, buyer context and trust signals with structured content, and measuring agent recommendations against an unenriched baseline. Dinh's advice stops short of a universal recipe precisely because the starting catalog varies, and because enrichment is not one size fits all.

1. Audit the catalog and classify it. Dinh asks merchants whether they hold a thin catalog or a spec-heavy one, because that answer determines which enrichment work has the most room to improve results.
2. Enrich with structured attributes rather than prose. Fill attributes, add description depth, add buyer context and trust signals, and add product identity data, following the pattern of the running shoe example.
3. Stripe boilerplate that repeats across listings. Welcome messages and store history add no retrieval signal and compete with the attributes that do.
4. Test against an unenriched baseline. PayPal's approach was to run experiments on agent search and recommendation rather than assume the outcome.
5. Recheck the balance between enrichment and focus. Dinh reported that over-enrichment raised hallucination in some cases, so more content is not automatically better.

The steps reproduce what PayPal described, and they apply to a merchant's own catalog rather than to a benchmark environment. Dinh noted that AI rewards content quality and optimizes for trust signals as well as brand presence. He also pointed readers to PayPal's Agent Commerce Services for more detail on the company's agentic commerce work.

FAQ

What is semantic search for AI agents?

It is retrieval that matches a shopper's request against the meaning of a product listing instead of its literal words, using vector representations of intent. Nixon Dinh of PayPal describes keyword search as precise but literal and semantic search as smart but fuzzy, which is why PayPal combines both in a hybrid approach.

Does enriching product data improve agent recommendations?

In PayPal's experiments it did. Dinh said enriched listings beat an unenriched baseline head-to-head, with thin catalogs and weak product data gaining the most room to improve. He reported the result as one set of experiments, not a general benchmark of all catalogs.

Can too much enrichment hurt?

Yes, according to Dinh. When merchants over-enriched listings with unstructured boilerplate, agents hallucinated more in some cases and recommendation performance dropped. Repeated store messaging competes with the attributes that actually carry retrieval signal.

Is hybrid search better than pure semantic search?

PayPal uses a hybrid approach that keeps keyword matching alongside semantic retrieval. Keyword search stays precise, fast and predictable for exact matches, while semantic retrieval handles sentence-length requests where the shopper's words never appear in the listing.

Why can't merchants just optimize catalogs for a single AI model?

Dinh described the model layer as largely a black box, with different agents using different models, ranking algorithms and non-deterministic vector retrieval. SEO offered a clear target, while agent optimization has no single standard yet, which is why he recommends running experiments.

Which merchants benefit most from catalog enrichment?

Merchants with the thinnest catalogs and the weakest product data, according to Dinh's summary of the experiments. A catalog that already carries rich spec data, materials and fit notes has less to add than one holding little more than a model number and a color.

What data should go into an enriched listing?

Dinh's example added a descriptive title, cushioning and breathability detail, attributes for fit and mesh material, and review data as a trust signal. Search could then match running, training and mesh rather than only the words in the original description.

Does brand strength substitute for catalog quality?

Dinh said AI rewards content quality, not brand, though it still optimizes for other trust signals. Reviews and structured product identity data give agents something to work with, while brand recognition alone does not change what retrieval can match.

What did PayPal project for agent commerce?

PayPal expects roughly 1.1 trillion US dollars of US retail commerce to be agentic and 3 to 5 trillion globally, and it cites a BCG projection that 15 to 25 percent of all commerce will be executed by agents by 2030. These are forecasts, not recorded results.

How much conversion lift did PayPal report from agent-first interactions?

Dinh said businesses that interact with an agent first tend to see up to four times higher conversion on order completion. He attributed the lift to the richer product data available inside that interaction, not to the agent itself.

From transcript to written article

The work behind this article was the same kind of step PayPal describes in its catalog advice: take material that exists in one form, give it structure, and let the quality of that structure decide whether anything downstream can use it. Dinh's warning about unreadable catalogs has a direct parallel in recorded expertise. A conference talk holds the evidence, the sequence and the caveats, but in audio form it stays out of reach of a reader or a search result.

Skala Blog turns a YouTube video into a written article through three steps: paste the video URL, let it produce a transcript, and generate the article from that transcript. The same logic applies to any recorded interview, lesson or product walkthrough in your own library.

If you have knowledge, explanations, interviews, opinions or lessons sitting inside a YouTube video, paste the URL at Skala Blog, let it transcribe the video, and turn that existing content into an article. Readers interested in the broader developer context behind tools like this can find the community at Crazystack, the home of Crazystack TypeScript and the Bootcamp do Dev Doido taught by Gustavo Dev Doido.

Source video