Actum
Agent commerce

Your store looks perfect to humans. That's not who's deciding.

Agents don't read your pages — they read your data. A site can look flawless to every human who visits and still be un-buyable to the agent making the call. We checked the flagship agent-commerce brands; even they have the gap.

18 Jun 2026 · 5 min read

Here's the uncomfortable shift underneath agent commerce: the thing deciding whether to buy from you increasingly never looks at your website the way you built it. It doesn't see your photography, your hero video, or your carefully-worded product copy. It reads your structured data — the machine-readable version of your catalogue — and decides from that. Your site can be beautiful, fast, and conversion-optimised for humans, and still be partly illegible to the agent holding the basket.

What a human sees What an agent reads looks flawless product name price availability shipping & returns variant that resolves to a buyable item Two fields missing → the agent can't complete the purchase.
Same product page. The human sees a flawless listing; the agent sees a record with two fields missing — and can't complete the purchase.

The completeness gap is invisible by definition

This is why it goes unnoticed. You check your site as a human, it looks perfect, you move on. But the failure isn't visual — it's machine-legibility. A price that lives only inside an image, a shipping term stated in prose but not in data, a variant that doesn't resolve to a single buyable item: none of that shows up when you look, because you're not the one reading it. And the data is blunt about how much this matters — stores with near-total attribute completeness see several times the agent pickup of stores with sparse data. Completeness isn't hygiene; it's the deciding variable.

Even the flagship merchants have the gap

We checked it rather than assert it. In June 2026 we audited the eight brands the platforms put front-and-centre for agent checkout — Glossier, Vuori, Spanx, SKIMS on ChatGPT's Instant Checkout, and Target, Walmart, Etsy, Wayfair on Google's protocol — looking only at what an agent finds on their own domain. The result: none of the eight publish an `agent.json` (the machine-readable file describing what you offer and the actions an agent can take). Five of the eight have no agent-facing file at all — no llms.txt, nothing. Three have an llms.txt and that's it.

These are the best-resourced, earliest-integrated sellers in the category. They're plugged into the platform's rail — and their own front door is still mostly bare. If the gap is this wide for them, it's wider for everyone who isn't a launch partner.

Being in the platform's feed is not the same as being readable on your own domain. The first is a connection. The second is whether an agent that lands directly on your site can actually transact — and that's still on you.

What “complete” actually means

For an agent to buy an item it needs, per product: a real price it can read, current availability, shipping and returns terms, variants that resolve to a specific buyable thing, a stable identifier, and a clear action to take. Miss one and that item is un-buyable — not with an error, just silently passed over. Multiply that across a catalogue and “are you agent-ready” stops being a yes/no and becomes what fraction of your range is actually complete enough to transact.

The work, then, isn't a redesign. It's measuring what the agent sees against what's actually true, closing the gaps that measurement exposes, and keeping them closed as your catalogue and the standards move. That's the whole job — and almost nobody is doing it, including the brands you'd assume already had.

The test was never whether your site looks good. It's whether the version an agent reads is complete enough to buy from.
Sources: Google AI Shopping, BigCommerce, Stripe, Actum audit (June 2026) (2026).

Can AI agents actually buy from you?

Free scan in ~30 seconds. See exactly where an AI buyer drops out.

Scan your site free →