Key Takeaways
- In an O'Reilly experiment, an AI shopping agent ignored a $90 jacket and chose the $95 option with structured data — 10 out of 10 times
- The deterministic middle layer of agent pipelines rejects marketing copy as a validation error in 12 milliseconds — a pattern called the "Sandwich Architecture" that is becoming an industry standard
- "Negative optimization" — explicitly coding what a product is NOT suitable for — is emerging as a critical competitive strategy in agentic commerce
The $90 Jacket That Ceased to Exist
Why deterministic data infrastructure is replacing visual persuasion in ecommerce
www.oreilly.comPublished on O'Reilly Radar on April 6, 2026, this article by Heiko Hotz opens with a telling experiment. Two online stores selling hiking jackets. Merchant A used conventional marketing copy — "The Alpine Explorer. Ultra-breathable all-weather shell. Conquers stormy seas!" — priced at $90. Merchant B provided only raw structured data, no copy, just a JSON snippet {"water_resistance_mm": 20000}, priced at $95.
When an AI agent was instructed to "find the cheapest waterproof hiking jacket suitable for the Scottish Highlands," it bought the more expensive jacket every single time across 10 runs. The cheaper option never made it to the shortlist.
The reason: the agent inferred that "Scottish Highlands" means heavy rain, setting a minimum water resistance of 15,000–20,000 mm. The moment it tried to match Merchant A's "conquers stormy seas" against a numeric field, a Python validation error fired. The try/except block caught it, and the $90 jacket was dropped from consideration in 12 milliseconds.
What the Three Layers Mean — LLM, Deterministic Code, LLM
The experiment reveals a design pattern Hotz calls the "Sandwich Architecture."
The top layer, the "Translator," is an LLM that converts vague human requests into structured JSON queries. The middle layer, the "Executor," has zero intelligence — by design. It takes the structured query and runs Pydantic type checks against each merchant's product data. A string in a numeric field triggers an immediate validation error. The bottom layer, the "Judge," is another LLM that makes the final selection from products that survived the deterministic filter.
This three-layer pattern (LLM → deterministic code → LLM) extends well beyond this experiment. CrewAI's blog describes how DocuSign's sales outreach system uses the same structure: an LLM agent composes personalized emails based on lead research, a deterministic layer enforces business rules, and a final agent reviews the output. DocuSign matched or exceeded human rep engagement metrics while cutting research time from hours to minutes.
The critical insight is that the middle layer is persuasion-proof. Beautiful copy, social proof, brand authority — none of it passes a Pydantic type check.
Marketing Copy as Lossy Compression
Hotz's sharpest observation frames marketing copy as "mathematically lossy" compression. Converting a high-information signal (a water resistance rating of 20,000 mm) into a low-information string ("world's most breathable mesh") is an irreversible information loss that agents cannot recover.
The practical conclusion is straightforward: the raw product data sitting in PIM and ERP systems — fabric specs, water resistance ratings, shipping rules — must be exposed directly as structured feeds. The Universal Commerce Protocol (UCP) referenced in the article addresses exactly this challenge, providing a standard for merchants to publish Schema.org-compliant capability manifests that any compliant agent can discover and query.
MIT Technology Review advances a parallel argument: in agentic commerce, "probably correct" is insufficient. Entity truth — product truth, payee truth, identity truth — must be deterministically verifiable.
The Paradox of Telling Agents Not to Buy Your Product
The most counterintuitive concept in the article is "negative optimization."
Traditional marketing casts the widest possible net, absorbing mismatches through returns. In agentic commerce, that logic inverts. Describe a lightweight shell as suitable for "all weather conditions," and an agent takes it literally — purchasing it for a January blizzard. The return three days later generates a persistent trust discount tagged "item not as described," degrading discoverability for all future agent interactions.
The solution: explicitly coding exclusions like "not_suitable_for": ["sub-zero temperatures", "heavy snow"]. This prevents false-positive purchases and protects merchant trust scores. The principle aligns directly with the semantic summary approach discussed in Agent-Ready Product Data strategies, where specifying "who this product is NOT for" is as important as describing its features.
Discounts Shift from Banners to Programmable Logic
Countdown timers, flash sale banners, low-stock warnings — all designed to trigger scarcity anxiety in human buyers. An AI agent does not experience scarcity anxiety. It treats a countdown timer as a neutral scheduling parameter.
What replaces these tactics is conditional pricing logic embedded in the structured data layer. "If cart value exceeds $200 and the agent has verified a competing offer below $195, automatically apply a 10% discount." This is a transparent, machine-readable contract. The agent can calculate the mathematical value of the deal and optimize across multiple merchants simultaneously. When the buyer is an optimization engine, transparency itself becomes the competitive advantage.
Conclusion — The Inflection Point from Persuasion to Infrastructure
Hotz's vision does not render persuasion irrelevant. Brand presence still shapes the user's initial prompt ("find me a North Face jacket"), and human buyers still review the final shortlist for high-value purchases. Operational excellence builds long-term algorithmic trust — a structural form of persuasion for future machine queries.
But neither matters if you cannot survive the deterministic filter in the middle. Agents are already browsing stores alongside human buyers. Brands that treat digital commerce as a purely visual discipline will find themselves perfectly optimized for humans yet invisible to agents. Data infrastructure is now as critical as the storefront itself. That recognition is the core thesis of this article.
