We launched Nohi in March 2026 with a small cohort of merchants who were willing to be early. That meant connecting their stores, letting us audit their catalogs, and trusting us to reformat and enrich their product data for AI agent retrieval.
It also meant we got a close look at what 18 real Shopify catalogs actually contain, and what AI agents actually do with them. These are our honest findings.
The catalog quality gap is bigger than we expected
Before we started, we assumed the main work would be technical: reformatting data structures, adding schema markup, syncing feeds. The actual bottleneck was something simpler and harder to fix.
Across the 18 catalogs, the median product description was 38 words. Most descriptions answered one question: "what is this?" Almost none answered the questions AI agents actually resolve.
Intent copy moved the needle faster than any technical fix
We tested two different optimization sequences. The intent-first merchants saw measurably better early results.
Before: "Stainless steel chef's knife. 8-inch blade. Full tang. Dishwasher safe."
After: "The everyday knife for someone who cooks four nights a week and doesn't want to think about maintenance. The full-tang construction stays balanced through repetitive prep work. Best for: home cooks, meal preppers, gift-giving. Not ideal for: professional-level butchering or precision Japanese-style cutting."
Catalog consistency amplifies individual product optimization
We noticed this pattern in several merchant catalogs where we had optimized the top 20% of products by sales volume, leaving the rest largely unchanged. The optimized products improved, but not as much as we projected.
Our working theory: AI agents that interact with a catalog make inferences about it as a whole.
Scope limitations paradoxically improved recommendation rates
We tested adding "not ideal for" signals to a subset of products. The expectation was that this might reduce recommendation rates. What we actually found was that recommendation rates went up.
The mechanism: when an agent has clear scope signals, it stops second-guessing whether to recommend a product for ambiguous queries.
