For twenty-five years e-commerce has optimized for the human end-user: fast UI, enticing photography, persuasive copy. In 2026 there is a new kind of customer, silent and structurally different: the AI agent. ChatGPT Shopping, Amazon Rufus, Perplexity Shopping, and the first MCP-native integrations on Claude and Gemini are shifting demand toward a funnel where the user delegates. The agent queries, compares, decides, purchases. And it judges your e-commerce by criteria your UX designer did not anticipate. This article explains what that means, and how to verify — today — whether your catalog is ready.
What an AI agent can and cannot do when it buys
A buying agent is a language model with three extra capabilities over a regular chatbot: access to external tools (MCP tools), memory of the conversation with the user, and an operational mandate (find, evaluate, buy within defined constraints).
What it does well: reads structured text, calls APIs, compares homogeneous data, respects explicit constraints («under 200 euros, delivery within three days, free returns»).
What it still cannot do: interpret ambiguous images with confidence, read narrative copy for hidden attributes, trust unstructured information. An agent that cannot find the return price in a readable format skips the seller. No asking, no inferring: it skips.
The Speed Trap
The first risk is what an emerging framework calls the Speed Trap: the agent has few tokens, little time, no patience. If your site needs JavaScript to load a price, if the page renders in two seconds instead of one, if the product page is spread across six sub-pages — the agent does not wait. It picks the competitor that responds first.
Speed Trap is not a UX-speed problem: it is an information-density-per-request problem. How much useful information can you hand an agent in a single query round? The difference between a ready catalog and an unready one sits here.
Machine Forgiveness
Second key concept: Machine Forgiveness, and its absence. A human forgives: if the product page is imperfect, they read reviews, ask the chatbot, call support. A machine does not. If the availability attribute is missing or malformed, the agent does not treat the product as «probably available»: it ignores it.
Zero Machine Forgiveness means data quality is not a nice-to-have: it is a threshold below which you exist or you do not. In practice: availability, price, returnPolicy, shippingDetails, brand, gtin, mpn must be present, correct, consistent across all channels (site, Google Merchant, Amazon, Meta Shops).
The Liability Vacuum
The third theme is more subtle but decisive in the long run: the Liability Vacuum. When a human makes a bad purchase, responsibility is theirs. When an agent does, whose fault is it? The model’s? The seller’s? The delegating end-user’s?
Jurisprudence is forming. But the behavior of the most conservative agents — OpenAI, Anthropic, Google — is already clear: preemptively skip sellers with ambiguous policies. If your return conditions live in a PDF linked from the footer, a liability-cautious agent avoids you. Not out of spite: out of design prudence.
The Comparability Bar
Fourth framework: the Comparability Bar. An agent always compares. It does not evaluate in isolation: it picks among alternatives. To compare, it needs attributes normalized across sellers. If you describe your product as «elegant and versatile» and the competitor says «100% organic cotton, 320 g/sqm, made in Portugal», the agent picks the second — even if yours is better.
Comparability means rewriting product pages with a spec-first logic: attribute first, narrative after. Culturally hard for brands built on storytelling — but the only way to compete in an agentic funnel.
How to verify readiness
Agentic readiness is measured on four axes with concrete metrics:
- Discovery — is your catalog queryable via MCP, structured feeds, llms.txt? How many agents read it today (GPTBot, ClaudeBot, PerplexityBot, Google-Extended)?
- Description — are product pages machine-first? Are Schema.org Product attributes complete and validated?
- Comparability — are your attributes normalized against market competitors? Are pricing, variants and availability declared coherently?
- Transactionability — can an agent checkout programmatically, handle idempotent errors, receive structured confirmations?
What to do, in order
Three interventions that move the needle in 60 days.
- Product schema audit: ensure every page has price, availability, brand, gtin, returnPolicy, shippingDetails, aggregateRating (if reviews exist). Validate with Google Rich Results Test and Schema.org validator.
- Machine-legible policies: move all policies — returns, warranties, shipping — into structured format. At minimum MerchantReturnPolicy and ShippingDeliveryTime. This alone moves your catalog from «risky» to «recommendable» for cautious agents.
- Catalog llms.txt: a /llms.txt file describing categories, brand tone, key policies. Not redundant with the sitemap: it is the narrative layer agents read to understand who you are before deciding to compare you.
How urgent is this?
The agentic commerce transition is not instant. Current volumes are still small — Bain and Forrester estimate 3–8% of e-commerce purchases agent-mediated by end of 2026. But the curve is exponential, and above all: first winners are building position now, not in two years.
Translated: if you are a mid-market e-commerce, investing in agentic readiness in the next six months is low-cost insurance with high upside. Not doing it is not an immediate risk — it is a silent erosion of visibility that you notice too late.