Agentic commerce: your catalog is now read by agents. Who writes it?
Meta Muse, Shop Pay checkout and NIQ's 51% make AI shopping agents a real channel. The product record is the new storefront, and agents write it too.
For twenty years, a wrong price on a product page was an embarrassment. A person saw it, maybe screenshotted it, maybe emailed support. In agentic commerce, a wrong price is an instruction. A shopping agent reads it, compares it, picks it and pays for it before anyone on your team opens the catalog.
That is the part of this month's agentic commerce news that sellers should read closely. The headlines are about who owns the checkout. The quieter change is about who reads your product data, and who writes it.
What changed this week
- Meta's Muse agent now takes shopping to checkout. At Connect, Mark Zuckerberg said Meta will make Muse "free for a huge number of tokens, with the expectation that over time we will profit by taking a small fee from transactions." New shopping connectors include Walmart, Best Buy, Sephora, Ulta and Wayfair, with Shop Pay and PayPal for payment.
- Shopify stores are reachable from Muse by default. Shop Pay checkout is live in Muse for Shopify merchants through Agentic Storefronts, which covers eligible US-selling stores without a separate integration. On its August earnings call, Shopify said agentic transactions carry the same economics as an online store transaction.
- More than half of US shoppers already use AI to shop. NielsenIQ's Agentic Commerce Tracker found that 51% of US consumers used at least one AI-powered tool to support their shopping in the past month, up from 42% in May. Product recommendations were the most common use.
- A category has formed around being chosen. Azoma, which calls itself an agentic commerce optimization platform, added Muse to the shopping agents it tracks, alongside ChatGPT, Gemini, Amazon Rufus and Walmart Sparky.
Taken together: an agent is becoming a real channel, the checkout is being wired straight into it, and a large share of shoppers already start there.
Your catalog is now a storefront nobody browses
When a shopper asks an agent for "a carry-on that works for two kids on a long flight", the agent does not scroll your homepage. It reads structured data: title, attributes, price, availability, reviews, shipping promise. Then it returns three answers, and the shopper usually picks one of them.
So the product page stops being the storefront. The product record is. Some people have started calling the work of getting picked agentic commerce optimization, the next layer after SEO and GEO. Whatever the name, it runs on the same raw material: product data that is complete, specific, and true.
Being complete and specific is a content problem. Staying true is harder, because of who is writing the catalog now.
The catalog is written by agents too
Look at how a small cross-border seller runs in 2026. One agent drafts listing copy for Amazon, Shopify and TikTok Shop. Another repriced last night against a competitor. A third pulled specs from a supplier PDF. A fourth updates stock from the warehouse sheet.
Every one of those is a write to the record that buying agents will read. And the failure mode is not the dramatic one. It is the plausible one:
- a price that is right for one channel, written to all of them;
- a battery spec copied from last year's model;
- "in stock" on a SKU that sold out an hour ago;
- a certification claim that no document supports.
Each passes every permission check. The agent was allowed to write the price. The value was well-formed. It was also false, and in agentic commerce the next reader is not a person who might notice. It is another company's agent, acting on it.
The buyer side already solved half of this
The payment side of agentic commerce has converged on one idea: do not give the agent the whole wallet. Stripe's Shared Payment Tokens give the merchant a credential scoped to one seller, capped by amount and time, and revocable. The agent never sees the card. In France, Cleverbridge, Visa and Revolut ran a live agent-initiated payment where Visa's Trusted Agent Protocol identified the agent and a passkey let the shopper approve the purchase. The agent gets the smallest authority that finishes the job.
The seller side needs the same shape, pointed at data instead of money. Not "can this agent write to the catalog", which is a yes, but "which writes should a person see before they become true":
| Write | Why it deserves a second look |
|---|---|
| Price changes beyond a threshold | Agents buy at the price you publish |
| New or changed spec claims without a source | A buying agent will repeat them as fact |
| Compliance and certification fields | A wrong "CE certified" is a liability, not a typo |
| Publishing a product to a new channel | It reaches every agent that reads that channel |
Everything else, such as copy tweaks, internal tags and draft descriptions, can stay fast. Review should be proportional to consequence. If every write needs approval, people start approving without reading, and then no write is really reviewed.
Three things to do before agent traffic matters to you
- Treat product data as the storefront. Audit the fields a buying agent actually uses (attributes, availability, price, shipping, returns) with the same care you give your homepage.
- Give every material field a source. A spec with a link to the datasheet can be checked. A spec without one is a claim, and it should wait for a person before it goes out.
- Separate who proposes from who decides. Let agents do the volume: drafting, repricing, syncing. Keep a person on the small set of writes that change what a customer pays or what you promise.
Where Busabase fits
Busabase is an open-source database and workspace for AI agents. Agents and people work on the same structured records, and writes that matter arrive as reviewable proposals with a field-level diff instead of landing silently.
In practice the loop runs in three places, and each has a free template you can install into a Busabase workspace:
| Step | Template | What your agents do there | What waits for a person |
|---|---|---|---|
| Get picked | Busa Agent Shelf | Ask the buyer questions you care about in ChatGPT, Gemini, Meta Muse and Rufus, record who gets recommended and why | Each proposed listing fix, with its source |
| Keep the record true | Busa Product Hub | Keep SKU data, prices, stock and channel status in one place, with an agent-readiness score per product | Price changes, channel publishes, unsourced spec claims |
| Close the order safely | Busa Agent Orders | Import agent-channel orders, flag anything paid below the listed price, propose repricing | Any price that crosses a guardrail, and how each exception is closed |
None of them replaces Shopify or your PIM. They sit where your agents write, so the record buying agents read is one a person has actually checked.
Sources: TechCrunch on Muse at Connect, Meta's Connect 2026 roundup, PYMNTS on Shop Pay in Muse, NIQ, Azoma, Stripe Shared Payment Tokens docs, Cleverbridge, Visa and Revolut. Verified 25 September 2026. This is a fast-moving area, so check the primary sources before you rely on any detail.