Chapter 9 — Be Usable

There is a lot of noise about agentic commerce and considerably less production reality. Both matter, and the gap between them is where budgets get wasted.

What’s actually happening

The forecasts are large. McKinsey projects that by 2030 the US retail market alone could see $900 billion to $1 trillion in “orchestrated revenue” from agentic commerce. Morgan Stanley projects $190 to $385 billion in US e-commerce spending from agentic shoppers by the same year.

Two things before either number goes on a slide. McKinsey’s “orchestrated revenue” counts any purchase journey an agent coordinates or influences — far broader than agent-completed checkouts. Morgan Stanley’s $385 billion is the top of a range and covers US e-commerce only. Placed side by side they look like independent confirmation of each other. They aren’t; they differ threefold mostly because they measure different things.

Meanwhile, in the actual market: in March 2026 OpenAI abandoned native in-ChatGPT checkout and refocused on product search and discovery, with purchases completing in retailer apps. The stated reason was that people research products in ChatGPT but don’t buy there, and merchant adoption was minimal — Shopify’s president said roughly a dozen Shopify merchants were using the AI checkout tools despite integrations being available across three platforms.

Independent measurement agrees. Of observed agentic browser activity, nearly 70 percent touched product and search pages. Just over 3 percent touched checkout.

AI commerce today influences discovery, not conversion. Budget accordingly. An “agentic commerce transformation” is funding a future. A clean product feed is funding the present.

The protocol map, briefly

Four acronyms circulate. Here’s what each is and whether it belongs on your roadmap.

UCP is the live one. Google announced it in January 2026, co-developed with Shopify, Etsy, Wayfair, Target and Walmart. By summer, Google Shopping’s Universal Cart had launched in the US across Search and the Gemini app, with UCP-powered checkout live at Nike, Sephora, Target, Ulta, Walmart, Wayfair and select Shopify merchants.

The single most concrete, checkable artifact in this entire space is UCP’s JSON manifest at /.well-known/ucp, which advertises your supported services and endpoints so agents can find your capabilities without a hard-coded integration. Your engineering team can verify in one HTTP request whether you have one. Very few brands do.

The eight attributes worth doing today

If you sell products, this is the most certain AI visibility work available, because it’s specified in public documentation rather than inferred from behavior.

Google connected eight Merchant Center attributes to AI Mode and AI Overviews, rolling out from January 2026:

Product Highlight — short selling benefits, 1–150 characters, two to a hundred per product. Google’s own language: helps customers discover products on AI-driven surfaces.

Product Detail — technical specifications by section, attribute and value.

Variant Option — differences beyond color and size.

Item Group Title — a generic title for a variant group.

Related Product — six relationship types, from “often bought with” to “substitute.”

Question and Answer — up to 1,000 characters each for question and answer, 10,000 combined per product, thirty pairs. Google says it’s “primarily intended for conversational experiences.”

Document Link — up to five PDFs such as manuals, which Google uses to answer detailed questions in AI Mode.

Popularity Rank — a 0–100 sales performance signal.

Note the irony in number six. Google fully deprecated FAQ rich results on websites during 2026. Meanwhile Q&A attributes in your product feed are explicitly built for conversational surfaces. The channel moved, not the content need. If you have a library of good customer questions sitting in deprecated FAQ blocks, they belong in the feed.

Can an agent actually finish a task on your site?

Beyond feeds and manifests, ask this once a quarter. Take an agentic browser, give it a realistic job — find a specific product variant, get a quote, check availability, download a spec sheet — and watch where it fails.

The failure modes are consistent and mundane. Content behind interaction: specifications in a tab that loads on click, pricing behind a configurator, availability behind a form with no URL state. Bot mitigation that challenges the agent and stops it. Unstable URLs, since agents re-fetch and session-dependent carts break on the second visit. Information that exists only inside images. Fulfilment terms written in prose that requires interpretation rather than in structured, machine-readable form.

Every one of those is also a human usability problem, which is the useful part. Agent-readiness work has a business case even if agentic commerce underdelivers.

Fund and defer

Fund now: a clean, complete, regularly refreshed product feed — the common denominator across every platform. The eight attributes, in the order listed. The /.well-known/ucp manifest if you sell where Universal Cart is live. Price, availability and fulfilment terms in the initial HTML. And a quarterly agent-completability test.

Defer: ACP checkout endpoints until OpenAI’s direction settles. And anything sold as an “agentic commerce readiness program” rather than a set of tickets. The work is a feed, a manifest, and a measurement loop. If a proposal is bigger than that, ask what specifically it ships.