PayPal Says Catalogs Aren't Ready for Agents

AI Engineer recently hosted PayPal director Nixon Den, who argued that keyword search is a dying breed. He believes catalogs built for humans will fail when AI agents start shopping based on intent.

The real issue isn't a lack of traffic, but a lack of readability.

Den broke commerce down into three eras. First is search, where shoppers typed "blue running shoes" and brands focused on SEO. Next is intent, where a shopper asks for marathon training support and agents use vectors to match the meaning. Finally, there is delegation, where agents are trusted to handle the purchase.

PayPal argues that most catalogs were designed for Google or Meta, not for agents. Since agents rely on different models and nondeterministic vector retrieval, there isn't a single ranking system for brands to game.

To test this, PayPal boosted product data with deeper descriptions, attribute filling, buyer context, and trust signals like reviews. This turned "men's shoe 8543" from simple blue running shoes into a lightweight cushioned mesh runner that matches queries for training and "true to size" fit.

In head-to-head tests, the enriched records won. Thin catalogs saw the most improvement, but adding unstructured boilerplate text diluted the signal, which sometimes pushed agents to hallucinate.

This retrieval fuzziness is why PayPal is co-developing a new transaction layer. In this system, the AP2 mandate carries the merchant identifier plus amount and expiry, allowing the merchant to verify the signature before processing the payment.

The lesson here isn't to add more text, but to provide structured quality that agents can actually reason over. Brand names alone won't be enough to win.