What Happens When AI Agents Start Choosing Which Products to Buy? (2026)

An intermediary may never watch your hero film. It matches constraints — size, compatibility, price cap, delivery window, in-stock, review evidence — against structured fields. Manifesto copy is not a spec. A worse product with a cleaner GTIN and honest availability wins.

Adobe Analytics, via Digital Commerce 360 (Aug 19, 2026), 1T+ U.S. retail visits, July 2026: AI-referral traffic +62% YoY, those visits converted 60% higher than non-AI traffic. That is humans arriving from generative tools — AI search as a channel. It is not proof that agents already complete purchase unattended.

The job. Make decision inputs honest — attributes, checkout-true price, fresh stock, delivery as data, versioned policy — so you enter the candidate set without lying.

This week. Open agent-choice-optimization-signals.md. Run 10 golden questions on one category. Fail any answer that asserts stock, price, or a discount without a tool/feed result.

A person still signs. Discount eligibility, cart assembly rules, and checkout capture. An agent that invents 15% off is ungoverned margin leak.

Skip it when required attributes are missing, stock is stale, or price on the feed disagrees with checkout. Brand campaigns will not fix a hollow catalog.

This is part 36 of eCommerce AI Agents. Executive readiness is your next customer may be an AI agent. Your copilot scorecard is post 10.

Our take: optimize decision inputs rather than a prompt you do not own. Brand storytelling gets less budget on the PDP that machines read. You remain a candidate when the intermediary never sees your hero video.

Copy the signals — agent-choice-optimization-signals.md. Channel score: agentic-commerce-merchant-readiness.md. Contract: ai-ready-catalog-contract.md.

FactualMinds is an AWS Select Tier Services Partner. We do not invent conversion lifts or GMV from agent-choice.

What’s already happening vs what’s still emerging

Same Adobe / Digital Commerce 360 article:

• AI-referral traffic +62% YoY, +1,219% versus October 2024
• Those visits converted 60% higher than non-AI traffic and generated 53% more revenue per visit (11th straight month of conversion outperformance in that series)
• +14% engagement, +59% time on site, −33% bounce, +28% add-to-cart versus non-AI (same article)
• Expanded cohort: 39% of homepages not machine-readable; apparel 76% LLM visibility

Intent mix bias is likely; do not copy 60% / 53% as a FactualMinds or campaign target.

Baymard 70.22% still describes what happens when totals, stock, and confidence break at checkout. An intermediary that chooses a SKU increases the volume of claims checkout will test.

What the intermediary weighs

flowchart TD
  req[Shopper requirements]
  attrs[Structured attributes]
  rank[Retrieval / ranking you do not own]
  rev[Reviews as data]
  price[Checkout-true price]
  avail[Honest availability]
  ship[Delivery promise as data]
  trust[Policy and identity]
  pick[Selected SKU]
  pay[Your checkout]
  req --> attrs
  attrs --> rank
  rank --> rev
  rev --> price
  price --> avail
  avail --> ship
  ship --> trust
  trust --> pick
  pick --> pay

GEO is not a guaranteed ranking technique. Packaging: post 34. You do not control the agent’s ranker. You control whether an honest SKU exists for it to pick.

There is no native Shopify AgentCore connector. External choosers will not wait for one.

FactualMinds helps retailers connect storefronts, catalogs, and back office so those inputs stay true — see retail / eCommerce on AWS.

What broke

What broke — Competitive “agentic merchandising” brief: increase brand adjectives on titles so assistants would “prefer us.” Width and waterproof remained unstructured; inventory asOf was 18 hours stale. Sampled queries picked a competitor with a worse review average and a complete spec. Detection: 10 golden questions — we lost on constraint match, not on storytelling. Fix: stop title stuffing; fill the catalog contract; fail in_stock when asOf exceeds SLA. Lesson: agents choose on fields. Copy is not a spec.

A second failure: issuing a one-off 15% code in the prompt so the “agent would win.” That is ungoverned discounting — same ban as the 15-automations map.

What to use instead

• A ranking vendor before the catalog contract. Score merchant channel readiness and the catalog contract first.
• Brand love copy instead of fields. Manifesto copy is not a spec.
• Payment-in-the-agent as week one. Checkout still owns card capture.

If you only do one thing

Run 10 golden questions on one category. If any answer asserts stock, price, or a discount without a tool/feed result, you are optimizing storytelling — not decision inputs.

For your technical lead

On June 17, 2026, AgentCore Harness reached general availability (What’s New). Useful when your agent compares SKUs with tools. Irrelevant as a magic ranking layer for everyone else’s agent.

AWS lifecycle notice (June 30, 2026) — Amazon Bedrock Agents Classic is in maintenance for new customers after July 30, 2026. If you host a shopping copilot, use Bedrock AgentCore. Full matrix: lifecycle roundup. Third-party agents that choose products do not run on your Classic stack.

First-party signals we reuse (not eCommerce outcomes) — Gateway server-side tools cut median tool round-trip ~180 ms → ~95 ms on a B2B CRM assistant (12 tools, ~8k turns/day) — Gateway post. Platform TCO silhouette: support-style AgentCore at 50K sessions/mo ~$791/mo platform + model (decision guide). Those numbers size your host. They do not measure agent-choice share of GMV. Model your mix on the AgentCore pricing calculator.

Gateway ~95 ms does not help if the offer is wrong. Your Harness copilot (GA June 17, 2026) can use the same contract. Skip Agents Classic after July 30, 2026.

What to do this week

1. Open agent-choice-optimization-signals.md. Run 10 golden questions on one category.
2. Fail any answer that asserts stock, price, or a discount without a tool/feed result.
3. Score merchant channel readiness and the catalog contract — do not start with a ranking vendor.
4. Pull AI-referrer metrics from post 33. Keep Adobe 60% / 53% as context, not an OKR.
5. Confirm checkout still owns card capture. AgentCore Payments GA is not permission to put PAN in a tool.
6. If you also want your copilot, use post 10 separately.
7. Contact FactualMinds, Generative AI on AWS, retail / eCommerce.

What this post doesn’t cover

• Autonomous shopper payment, card-on-file, or unbounded x402 checkout — AgentCore Payments GA (August 18, 2026) covers gated agent-to-API spend, not replacing your processor. Architecture: AgentCore production guide.
• How to rank inside a named third-party assistant
• Measured FactualMinds GMV from agent-choice — we are not inventing it
• On-site copilot 13-check list — post 10
• PIM extract/publish — post 9
• Speculative AGI shoppers

It does not invent conversion lifts. +62%, 60%, 53%, 39%, 76%, 70.22%, ~180→95 ms, and ~$791/mo are the published figures we reuse.

Primary next step: contact us or see Generative AI on AWS and retail / eCommerce.

FAQ

When should you NOT optimize for AI agents choosing products?

Skip a dedicated program when required attributes are missing, stock is stale, or price on the feed disagrees with checkout. You are not in an honest comparison set. Brand campaigns will not fix a hollow catalog. Score the merchant-readiness and catalog-contract artifacts first. Do not invent a 15% code so an agent can “win” the session.

What could go wrong if you treat Adobe’s AI-referral conversion lift as proof agents already buy for shoppers?

Adobe Analytics (July 2026, 1T+ visits) measures humans clicking through from generative tools: +62% YoY traffic, 60% higher conversion, 53% more revenue per visit, +1,219% vs October 2024. That is current-channel mix, with likely intent bias. Collapsing it with autonomous merchant selection is how you staff payment-in-the-agent and skip the catalog contract. Instrument referrals; optimize decision inputs separately.

What could go wrong if merchandising optimizes “brand love” copy instead of decision signals?

Agents match constraints: size, compatibility, price cap, delivery window, in-stock, review evidence. Manifesto copy is not a spec. You lose to a worse product with a cleaner GTIN, parent/child graph, and honest availability. Trust signals that matter are versioned policy and identity — not a slogan.

Do we control how a third-party agent ranks us?

No. You control whether you enter the candidate set with true attributes, offer, and policy. Ranking inside ChatGPT, Perplexity, or a shopping agent is their system. GEO is not a guaranteed ranking technique. Accurate representation is the job; position is not for sale as a FactualMinds KPI.

Should we let an agent invent a discount to win merchant selection?

No. Discount eligibility is an OMS rule. An agent that invents 15% off to beat a competitor is ungoverned margin leak — the same failure as a dishonest on-site copilot. Assemble a cart; hand off to checkout. Cedar belongs on any write your own agent can call.

Is AgentCore required for agents to choose our products?

No. External agents need your catalog contract. AgentCore Harness (GA June 17, 2026) is how you host YOUR copilot or draft loop. Skip Agents Classic after July 30, 2026. Do not buy an agent runtime to fix missing width or stale stock.

Need the decision inputs — catalog, offer, policy — honest enough that an intermediary can choose you without lying? Contact FactualMinds or see Generative AI on AWS and retail / eCommerce.