Every briefing so far has argued that a store an agent can read gets recommended and one it can't gets skipped. This cycle someone stopped arguing and measured it. An audit scored 141 product pages at 29 leading retailers against the requirements agents actually read. Seventy percent fail, and they mostly fail on the same three fields.
For three briefings running, the claim has been the same. AI agents are doing more of the shopping, and the ones that can't read your store simply move on to one they can. It was a reasonable claim, backed by traffic numbers and vendor surveys, but nobody had walked up to real stores with the actual checklist and marked them.
Now somebody has. An independent audit pulled 141 product detail pages from 29 leading retailers and scored each one against Google's Universal Commerce Protocol, the specification that says what an agent needs to find on a page before it can transact. Seventy percent came back as failing. Not thin, not unpolished. Failing.
The basics were fine. It was the fields that let an agent finish the purchase that were missing.
Agents don't judge your store on the parts you spent money on.
A person landing on your product page reads the photograph, the copy, the reviews, and a price. An agent reads none of that the way you intend it. It goes looking for a small set of stated facts, published in a form a machine can parse without guessing: what the product is, what it costs, whether that price is still good, how long it takes to arrive, how many days a buyer has to send it back, and the standard product number that lets it match your item against the same item elsewhere.
Those are not exotic requirements. They are the things a buyer asks before handing over money, written down so software can read them instead of inferring them. What the audit found is that almost every store publishes the easy half and almost none publish the half that closes a sale. Price was there on 99 percent of pages. Brand on 96 percent. The delivery and returns facts an agent needs to commit were on roughly one page in eight.
Last cycle the checkout went live. This cycle the shelves got marked, and buying crossed into B2B.
In our last briefing the news was that agents could finally pay: the rails became a shared standard and the first live agentic card payment settled on a real merchant. That removed the last human from the loop. What it left open was how ready anyone's catalog actually was.
The audit answers that, and the answer is worse than the earlier estimates suggested. It also turned up something blunter than a missing field. Fifteen percent of the pages it tried returned an HTTP 403 and refused the crawler outright. Those stores are not scoring badly on readability. They are not in the room.
Share of 141 audited product pages carrying each machine-readable field. The facts that identify a product are nearly universal. The facts that let an agent commit to a purchase are not.
Source: SALT.agency audit of 141 product detail pages across 29 leading retailers, scored against Google's Universal Commerce Protocol requirements, published July 21, 2026. A single agency's study, with its method and per-field results public.
The second change is who is buying. Agentic purchasing stopped being a consumer story this cycle. Visa and the payments firm Lianlian completed Greater China's first live business-to-business agentic transaction, where the agent identified a purchasing need, compared suppliers, placed the order and paid, all inside the company's spending controls. Days earlier Salesforce shipped a business commerce release with a buying agent that completes purchases over WhatsApp and SMS with no sales rep involved.
Being findable was never the test. Being buyable is.
Think about what an agent does with a missing field. It has a shopper's instruction, a budget, and a deadline. It finds your product and cannot confirm the price is still good, cannot say when it lands, and cannot tell whether it can be sent back. It does not email you to ask. It moves to the store that published all three and buys there, and nobody ever tells you that you were in the running.
That is not a hypothetical audience. A survey published this cycle found 74 percent of consumers have already used an AI agent while shopping, with comparing brands and finding where to buy as the top two uses, both of them discovery jobs that need a readable catalog. On the size of it, WARC and Omnicom's PHD put consumer spending that AI agents have a hand in at 944 billion dollars this year, rising to 3.35 trillion by 2030, close to four percent of all global consumer spend. The share of your category moving through agents is only going one direction.
And if you are one of the fifteen percent returning a 403, none of the rest applies yet. You can have every field published correctly and still lose, because the agent never got through the door to read them.
A separate, related free tool
The 403 problem is usually an accident of configuration, not a decision. Our free AI Visibility Checker writes you a correct robots.txt so the agents you want can reach your site and the ones you don't stay out. It matters most if you're on Cloudflare, whose defaults change on September 15, 2026.
Set your AI crawler rules →No sign-up, about two minutes, and it doesn't touch your search ranking.
You already hold every field the audit says is missing. It just isn't anywhere an agent can ask for it.
Read that list of failures again and notice what it is not. It is not a data problem. Your order system knows the delivery window. Your returns policy has a number of days in it. Your inventory system carries the product ID. Every fact the audit marked absent already exists inside your business, usually behind an API some internal tool talks to every day. It is simply not published anywhere a machine shopping on a customer's behalf can reach it.
That gap is the whole of what we build. BeaconSpec turns the REST and GraphQL APIs you already run into one curated server that AI agents can discover and buy through over the open UCP standard, with the shopper signed into your own accounts. Instead of hoping an agent scrapes the right facts off a page, it asks your systems directly and gets the real answer, live. You choose exactly which operations are exposed, and you can host it yourself if your rules require that.
This is also where the split in the market shows. If you run on a large commerce platform, someone is building this path for you. If you run a custom or home-grown stack, and most businesses selling to other businesses do, nobody is, and you are the store scoring low or not appearing at all. Worth knowing: a retailer that started this cycle went from concept to a production AI shopping assistant in about six weeks. This is not a multi-year program. It is a translation job on data you already have.
The audit found missing fields. What it really found is businesses whose systems know the answer and have no way to say it.
Agents don't reject you on price. They skip you over three fields you never thought to publish.
Price validity, delivery time, and return window appear on roughly one product page in eight. Without them an agent cannot commit, so it buys from whoever published them. Your systems already hold all three.
This is a market briefing we run every cycle, tracking what's actually changing in agentic commerce so you don't have to piece it together yourself.
The rule has not moved since the first briefing: readable gets recommended, unreadable gets skipped. What each cycle adds is precision. First we knew agents were choosing. Then we knew they could pay. Now we know exactly which fields they check, and that seven in ten of the biggest stores in the market do not publish them. That is unusually good news, because a gap this specific is a gap you can close, and right now most of your competitors have not.
BeaconSpec exists for exactly this shift: making your business discoverable and transactable by AI agents over the UCP standard, instead of invisible to them.
See what we do • Read more articles
Related reading: The money went live: agents can now buy, if they can read you
David Soden writes about agentic commerce, automation, and building durable technical systems for businesses. Photography via Pexels (iMin Technology, Tima Miroshnichenko, GB The Green Brand); figures cited are drawn from the public reporting named above.