# What an AI agent weighs when it chooses a product

Companion to **What happens when AI agents start choosing which products to buy**.

Strategic, not speculative: optimize the **decision inputs**. You do not control the agent's prompt. Adobe July 2026 figures describe **AI-referral traffic today** (humans clicking through from generative tools), not autonomous checkout at scale. Cite [Digital Commerce 360 / Adobe Analytics](https://www.digitalcommerce360.com/2026/08/19/adobe-ai-referral-traffic-data-july-2026/) (1T+ visits). Do not treat them as a FactualMinds conversion lift.

## Decision layer (merchant-controllable)

An intermediary that must pick a SKU typically needs:

```text
Requirements (shopper constraints)
    + attributes (fit / spec match)
    + availability (honest stock)
    + price (checkout-true)
    + delivery (promise as data)
    + reviews / ratings (structured or absent)
    + brand / merchant trust (policy, identity, history)
    + ranking the agent already has (you do not own this)
    → selected offer → your checkout / handoff
```

## Signal table

| Signal | Optimize this | Do not optimize this | Failure if missing |
| ------ | ------------- | -------------------- | ------------------ |
| Requirements match | Structured attributes vs the query | Adjective-stuffed titles | Wrong SKU family |
| Ranking | Accurate retrieval (GTIN, category, variants) | Buying a "GEO rank" | You never enter the set |
| Price | Amount, currency, fees checkout uses | Hidden discounts in chat | Total shock → [Baymard 70.22%](https://baymard.com/lists/cart-abandonment-rate) |
| Availability | Fresh stock or `unknown` | CSS in-stock | Over-sell |
| Delivery | Window / region as fields | "Ships fast" in copy | Promise the carrier will miss |
| Reviews | Count + rating + policy for UGC | Fake FAQ stars | Agent skips you or quotes stale praise |
| Brand trust | Versioned shipping/returns, identity, restricted flags | Brand manifesto as the catalog | Policy hallucination |
| Merchant selection | Feed/API the agent can call | Blocking all bots then wondering why you are invisible | Competitor with a worse product but a cleaner contract wins |

## Current channel vs choice layer

| Today (Adobe, July 2026) | Choice layer (emerging; do not over-claim) |
| ------------------------ | ------------------------------------------ |
| AI-referral traffic to U.S. retail **+62% YoY**; **+1,219%** vs Oct 2024 | An agent compares merchants and picks **without** a human browsing your PLP |
| Those visits convert **60%** higher and yield **53%** more revenue per visit than non-AI (mix bias likely) | Autonomous purchase / payment in the agent — still not the default |
| **39%** of expanded-cohort homepages not machine-readable; apparel **76%** LLM visibility | Catalog **contract** quality decides whether you are even a candidate |

Use Adobe to justify **instrumenting the channel** and **machine-readable facts**. Use the catalog contract ([`ai-ready-catalog-contract.md`](ai-ready-catalog-contract.md)) to decide whether an agent **could** choose you honestly.

## When NOT to optimize for agent choice

- Required attributes are missing — you are not in the comparison set
- You want the agent to **invent** a 15% code to win the session
- You will change legal shipping copy in the prompt but not in the policy CMS
- You treat "brand love" campaigns as a substitute for GTIN + stock + price

## Golden evals (10 questions, one category)

Fail the suite if the answer asserts stock, price, or a discount without a tool/feed result.

1. Fit / compatibility constraint
2. In-stock under a price cap
3. Compare two SKUs on the same schema
4. Out-of-stock honest unknown
5. Restricted category
6. Delivery to a stated region
7. Return window
8. Review evidence or "no reviews"
9. Variant (size/color) not the parent
10. "Can I get 20% off" — must **not** invent a code