BeaconSpec

← All articles

Agentic commerce • Field briefing

Google wired thirteen apps into Gemini. There was no application form.

A billion people now use Gemini every month, and this month thirteen outside companies got a door into it. If you sell in one of those categories and you weren't on the list, nobody called you, and there is nowhere to apply.

David Soden  •  Agentic commerce  •  6 min read
A red carpet running between steel posts and rope barriers on a grey tiled plaza, an entrance open only to people on the list.
Thirteen companies were shown in this month. The rope is the interesting part, not the carpet.

On August 12 Google said two things in the same breath. Gemini had passed a billion people using it every month, and thirteen outside apps can now be used from inside it. Angi and Thumbtack for home professionals. Zocdoc for doctor appointments. Ticketmaster, Fever and GetYourGuide for events. OpenTable for restaurant tables in the UK, Localiza for rental cars, plus Wix, Otter.ai, Granola, iHeartRadio and Pandora.

Read the category list again, because this is not a retail story. Home services, healthcare, events, dining, car rental. Thirteen companies got a door into the assistant a billion people already carry around. Everyone else in those same categories got nothing, and there was no application window, no waiting list, no form. If you run a clinic, a hotel group, a restaurant chain or a home services business, your path into that assistant this month was to be one of thirteen phone calls Google decided to make.

Being reachable by an AI right now is an invitation list, and you are not holding the pen.

What's happening

Reachability is being handed out, not earned.

Assistant makers are filling categories by partnership because it is fast and they stay in control of the result. That is a rational thing for them to do and a bad thing to depend on. Look at how quickly the rest of the stack got claimed: Stripe agreed on August 16 to pay more than $7 billion for OpenRouter, the gateway that routes traffic between AI models, and on August 13 the Linux Foundation's Agentic AI Foundation, the neutral home of the Model Context Protocol, announced it had reached 247 members, with Visa, Alibaba and Wells Fargo joining at its top tier.

Models, payments, governance, distribution. Each of those layers now has somebody large standing on it. One layer still has nobody standing on it, and it is the one that belongs to you: the interface a business publishes about itself, what it sells, what an agent may do, and whose account the agent is acting in. No platform vendor is going to ship that to a business running its own booking engine or its own order system.

Notice the word the industry settled on while all of this was happening. OpenAI and AWS published their pattern for agents that pay on August 13, and they called it "Controlled Agentic Commerce". Control was the point. It has to exist on your side of the wire too, or you are just hoping.

What's changed since our last briefing

Two weeks ago you couldn't keep agents out. This week you can't get yourself in.

Our last briefing covered a federal appeals court throwing out Amazon's block on Perplexity's shopping agent, on the reasoning that when an agent shops for a person, it is the person doing the accessing. Blocking stopped being a plan. That was the defensive half of the picture.

This is the other half, and it is the more uncomfortable one. You cannot keep agents out, and you cannot get in by asking either. Both halves land on the same lever: what you publish. Hotels are hearing it in their own trade press this month, where Hospitality Net ran a piece headlined "Most Hotels Are Invisible to AI" alongside the finding that 52% of UK travellers now plan trips with AI, and Agoda's data puts 63% of active travellers in Asia Pacific using AI to plan. Radisson wired hotel search into ChatGPT on August 17. The businesses next door are starting to move.

A signpost photographed from behind against a blue sky, every direction blade blank with nothing written on it.
This is what an agent finds at most businesses today: the post is there, the signs say nothing it can read.

Almost nothing on the shelf clears the bar

Product-page readiness scored against what an AI needs before it will recommend you at all.

48 61 <1% Average page scores 48 out of 100 Score needed to compete: 61 Pages that clear the minimum bar

Source: Mirakl audit of 427 product pages across 35 countries, published August 2026. The same audit found 43% of pages carry no reviews or ratings and only 9% publish machine-readable product data. Mirakl sells a fix for this, so treat the exact figures as directional; the methodology is stated and the direction matches everyone else's.

Why this matters to you

The buyers are arriving at doors that were never built.

Arkose Labs, which watches bot and agent traffic for its customers, measured agentic traffic across that base up 1,100% from the first quarter of this year to the second, and expects another spike over the holidays. That is one vendor's own telemetry rather than a market measurement, so hold the exact number loosely, but the order of magnitude turns up everywhere. ICSC and McKinsey put US agentic commerce revenue at a trillion dollars by 2030, with 43% of consumers already saying they trust an agent for simple purchases. Gartner's figure for agent-intermediated B2B purchasing is $15 trillion by 2028.

Now put those two facts next to each other. Agent visits are multiplying, and under one in a hundred product pages carries enough for an agent to act on. The gap does not arrive as a complaint. It arrives as nothing: no enquiry, no abandoned cart, no bounce you can point at in a report. The agent read what it could find, decided it could not commit, and recommended a business it could. VentureBeat spent August 11 on exactly this, why AI-driven purchase intent so rarely turns into a completed sale, and landed on the handoff: intent gets created somewhere you do not own, and there is no clean path from there into an actual order.

There are only two ways an agent finds you

One of them needs somebody at Google to think of you first.

THE INVITATION LIST (TODAY) Your business Wait to be picked Thirteen seats, filled no form, no window, no phone number A PUBLISHED INTERFACE The APIs you already run One curated card you publish Any agent can find you and buy nobody has to pick you first The top path depends on a call you cannot make happen. The bottom one you can start this quarter.

Why we're built for this

A published interface is the only door that does not need permission.

BeaconSpec is the discovery layer for agentic commerce, and this week is the whole reason it exists. Point us at the REST, OpenAPI or GraphQL APIs your business already runs. We turn them into one curated server an agent can call, and publish it as a UCP server card at your own /.well-known, so ChatGPT, Gemini or Perplexity can discover it without anybody at those companies choosing you first. UCP is the open standard Google, Shopify, Walmart and Target are building on, and MCP, the protocol underneath it, is the one that just landed under a foundation with 247 members and a card network at its top tier.

Two parts of that matter more than the plumbing. The first is curation: you decide exactly which operations an agent can ever see, and nothing else is exposed. An open endpoint over your whole API is a liability, not a strategy. The second is the sign-in, because real commerce needs the shopper actually logged in before "where is my order" or "place my order" means anything. We handle the standard login handshake so the agent acts inside your customer's own account, with no shared passwords and no separate build per AI vendor. Run it hosted by us, or take a sealed container and run it yourself if your data cannot leave your walls. Keeping up with the specs as they move is our job rather than a standing item for your engineers.

A smaller, separate thing worth doing

Nothing to do with the invitation list, but easy to get wrong: your own site settings may already be turning AI crawlers away, and on September 15 Cloudflare's default flips for sites that never set their own rules. Answer three plain questions and our AI Visibility Checker writes a correct robots.txt for you.

Check your AI visibility →

No sign-up. Copy the result and paste it into your site.

A hand resting on a metal service bell at a dark hotel reception counter, a front desk anyone can walk up to and ring.
The opposite of a guest list is a front desk: a published way in that works for whoever shows up, including a machine.

The one thing to remember

Thirteen companies were let in. Nobody is coming to let you in.

A billion-user assistant filled home services, healthcare, events and dining by picking partners, and there is no application process for the businesses it skipped. Agent traffic is multiplying while under 1% of product pages carry enough for an agent to act on.

The only door that does not need permission is the one you publish yourself: a curated, standards-compliant interface any agent can discover, use and buy through, with you deciding what it can see and do.

The thing worth watching now is whether Google replaces the hand-picked partner model with an open protocol path. If it does, the businesses that already publish a standard interface get in on day one and everyone else starts a project. If it does not, the invitation list simply carries on, and it stays short. Either way the answer for a business that is not on it is the same, and it does not involve waiting for the phone to ring.

BeaconSpec exists for exactly this shift: making your business discoverable and transactable by AI agents over UCP, on an interface you publish and control, instead of waiting to be picked.

See what we do  •  Try the free AI Visibility Checker  •  Read more articles


David Soden writes about automation, web strategy, and building durable technical systems for businesses. Photography via Pexels (Jan van der Wolf, Mikhail Nilov). Figures cited are drawn from Google's own announcement of Gemini's connected apps (August 12, 2026), the Linux Foundation's Agentic AI Foundation membership release (August 13, 2026), Mirakl's product-page readiness audit (August 2026), Arkose Labs' agentic traffic report (August 11, 2026), ICSC with McKinsey's "Shopping in the Age of AI", and Gartner's B2B agent forecast.