Specoria
Menu
TR Free test

InsightsAgentic commerce

Seven things people get wrong about agentic commerce

Agentic commerce isn't just a buy button in ChatGPT, llms.txt isn't a ranking factor and no single protocol will win. Seven myths, the sources, and what to do.

By 8 min read
A cut-paper browser frame stands on a beige background with blank tags, small parcels and a torn paper arch around it, and a single violet ribbon.
The shortlist opens the door long before the checkout does. Illustration: Specoria · generated with AI

Ask five store owners what "agentic commerce" means and you will hear five different answers. One pictures a buy button inside ChatGPT. Another thinks it is a file you add to your website. A third has decided it is a problem for big US retailers and can wait until next year.

Each of these ideas leads to spending in the wrong order: a checkout integration before the product data is right, a new file before the existing pages are readable, or nothing at all while competitors learn. In this article, we go through seven common misreadings, what the primary sources actually say, and what to do instead. Our own view runs through all seven: data before protocols, discovery before checkout.

1. "Agentic commerce means checkout inside ChatGPT"

The best-known example of an agent buying on a shopper's behalf was OpenAI's Instant Checkout. In March 2026, OpenAI changed course. A spokesperson told Digital Commerce 360: "We're prioritizing making ChatGPT search and product discovery great, with ACP serving as the infrastructure that connects users to merchants across the full shopping journey. Instant Checkout is moving to Apps, where purchases can happen more seamlessly."

Shopper behavior points the same way. In Forrester's December 2025 survey, 23% of Gen X, 32% of Millennial and 35% of Gen Z US online adults had used ChatGPT in the past month to search for products. Forrester is explicit that this is browsing, not purchasing.

So today, most of the value sits earlier in the journey: being found, compared and recommended. Checkout in an AI surface exists and will grow, but a store that is never on the shortlist never reaches it.

What to do: treat recommendation as the first goal. Make sure an agent can read your products, prices, stock and policies before you plan any in-chat checkout. We covered how agents build that shortlist in How do shopping agents choose a store?

2. "llms.txt is a ranking factor"

A stack of blank cards sits in the center with four thin paper bridges fanning out to an envelope, a parcel, a key and a blank card, the key in violet.
One clean source of truth can feed every protocol that arrives later. Illustration: Specoria · generated with AI

llms.txt is a proposal, written by Jeremy Howard in 2024 and updated as v2 in August 2026, for a Markdown file that gives agents a short guide to a website. The proposal itself notes real adoption: thousands of sites publish one, and Chrome's Lighthouse checks for it in its agentic browsing audits. That makes it a useful convention.

It does not make it a ranking signal. Google's guidance for AI Overviews and AI Mode says there are "no additional requirements" to appear in these features, and states: "You don't need to create new machine readable files, AI text files, or markup to appear in these features."

What to do: publish an llms.txt if it is cheap for you, and keep it accurate. Think of it as good hygiene, like a clean sitemap. Then spend the real effort on the pages and feeds it points to.

3. "We have an API, so we're ready"

An API is a good start, but an agent needs more than an endpoint. It has to find the API, understand what each call does from the documentation, get permission that is limited to what the shopper approved, retry a request without creating a second order, and get answers fast enough to keep the conversation moving.

There is also a simpler problem. An API returns the same catalog data your website shows. If a variant's stock is wrong on the page, the API delivers the same wrong answer, only faster.

What to do: before you call the API agent-ready, check five things: is it discoverable, documented, scoped, safe to retry (idempotent) and fast? Then check that the data behind it matches your page and your feed.

4. "Agents only compare price"

Price matters, but the formats built for agents ask for much more. OpenAI's product feed specification for ChatGPT lists optional fields for variants, product attributes, shipping, returns and reviews next to the required price and availability. In the Universal Commerce Protocol (UCP) catalog example, each variant names its seller and links to that seller's refund policy.

When two stores sell the same item at a similar price, the details decide: which one ships sooner, which one states its return window clearly, which one shows stock for the exact size the shopper asked for.

What to do: make delivery times, shipping costs and return terms as easy to read as your price. Our guide on how shopping agents read your return policy shows how to turn policy text into numbers an agent can compare.

5. "This is only a retail thing"

Retail is where most of the attention is, but it is not where the standards stop. UCP's own site says the protocol is "expanding beyond retail": a lodging draft specification is available and food is listed as coming soon.

On the business side, Gartner predicts that by 2028, 90% of B2B buying will be AI agent intermediated, pushing over $15 trillion of B2B spend through AI agent exchanges. That is a forecast, not a measurement, but it shows where analysts expect the same pattern to spread.

What to do: if you sell services, rooms or to other businesses, the basics are the same. Publish prices, availability and terms in a form a machine can read, and keep them consistent everywhere they appear.

6. "Shoppers don't trust agents yet, so we can wait"

Trust is moving faster than many expected. Adobe reports that in March 2026, AI traffic to US retail sites converted 42% better than non-AI traffic. In March 2025, it converted 38% worse. That is US data, and conversion from AI referrals is not the same as purchases made by agents, but the direction is clear: shoppers who arrive from an AI tool increasingly arrive ready to buy.

Waiting has a cost that is easy to miss. The stores that start now learn which products agents pick up, where they get stuck and which fixes help. Stores that wait start that learning later, against competitors who already have it.

What to do: start measuring now, even with simple tools. Our article on measuring the agent channel explains why one screenshot of an AI answer is not enough and what to track instead.

7. "One protocol will win, so we'll bet on it"

The protocols people compare are mostly solving different problems. The Agentic Commerce Protocol (ACP), created by Stripe, OpenAI and Meta, standardizes checkout between agents and businesses. UCP, launched by Google in January 2026, covers the journey from catalog search to orders. Google donated the Agent Payments Protocol (AP2) to the FIDO Alliance in April 2026, and on April 2, 2026, the Linux Foundation announced the x402 Foundation to steward the x402 payment protocol. Model Context Protocol (MCP) is an open standard for connecting AI applications to data sources and tools.

These are layers, not rivals in a single race. A store may end up using more than one, and the list will change.

What to do: keep one source of truth for product, price, stock and policy data, and treat each protocol as an adapter on top of it. When a new standard arrives, you add an adapter instead of rebuilding the store. Our agent protocols overview shows what each one is for.

The bottom line

Most of these myths share one mistake: they put the visible, new piece (a checkout button, a file, a protocol) ahead of the plain work underneath it. Agents recommend and buy from stores whose data they can read and trust. That foundation pays off on every surface, whichever protocol ends up carrying the order.

Specoria measures that foundation: whether an agent can read your product data, find consistent prices and policies, and get from product page to checkout. You can see where your store stands with a free readiness test.

Sources