Last week we pointed a small program at 189 industrial websites and asked each one something a purchasing clerk asks all day. Does this part exist, what does it cost, when can you ship it.
The program does not open a browser. It sends a request the way a buying agent sends one and writes down what comes back.
94 of those 189 sites had nothing to say to it.
Try one yourself. Pick a distributor you buy from and open any product page. Everything is there: part number, pack size, your price, a stock indicator. Now ask the same page the same question from a script, with no browser and no login, and a lot of these sites answer 403. Same company, same part, same catalogue. The only thing that changed is who is asking.
You will not find that in a dashboard. Dashboards are built from people opening pages, and no person's browser ever trips it.
It is closer to sediment than to a decision. These sites were built for a person with a mouse. Then, for twenty years, the only programs that asked them anything were scrapers, so the sites got hardened against programs. A firewall cannot tell a scraper from an agent buying on behalf of your customer, so it does the cautious thing and blocks both.
The catalogue itself lives in an ERP. What reaches the web is a rendering of it, a page that reads like a catalogue to you and like furniture to a machine. One product page we scanned returned 384 characters of text. The other thirty thousand characters were navigation, a cookie notice and script.
Now sit at a quote desk for a minute. A request arrives by email. Someone opens it, reads it, retypes it into the system, sends a quote back the next day. That has been the job for decades, and it is a real job with real people in it. A buying agent does the same work in about four seconds and never gets tired. When it cannot read you, it does not make a note to try harder. It goes to the next supplier. You do not lose an argument. You are just not in the comparison.
The fix is smaller than it sounds, and it is not a new website or a new ERP.
It is a service window standing in front of the systems you already run. An assistant asks a question in one of the six forms a clerk already understands: does this item exist under this number, what is the price for this buyer, is it in stock, when can it ship, what replaces it, what is the minimum order. Everything else is refused until you decide otherwise, and the refusal is a named answer rather than silence.
Your rules stay yours. Which buyer sees contract pricing, which sees list pricing, which sees nothing at all. The gateway answers out of your ERP and your price list, and for the first month it is read-only: it can quote, it cannot take an order. Nothing is copied out of your building. The log of who asked what comes back to you rather than sitting with us.
The difference you notice first is not extra traffic. It is the requests you never see today. An agent gets 403 and leaves, and nobody learns it was a buyer, or which product line it wanted, or that it compared you against four others in the same minute. Readable turns that visit into a question you can answer.
There is a second half, and when it is missing, it is the part that ends up in a dispute. An agent asks for a quote and gets one, and then the supplier's version of what was said and the buyer's version drift apart, with nothing in the middle to settle it three weeks later. So we published a rule, free to use and free to check: a signed receipt on each step. A line a person can read, a fingerprint of what moved, both sides named, a signature from each, the time. Change one character and the receipt stops matching. Spend the same approval twice and it is refused.
What a receipt proves is narrow and we say so on the page rather than burying it. It proves both sides signed that line and that nothing changed afterwards. It does not prove the line is true, and it says nothing about whether the goods are any good. No receipt is not a clean step either. It is an unknown. Your buyers, brokers and auditors verify the receipts themselves with keys we publish, and it costs them nothing.
We applied the same rule to ourselves before asking anyone else to. Our own machine-readable card states what our agents may do and what stops them.
Almost every rule that governs this was settled in the last year. How an assistant asks your catalogue a question became a standard called MCP, published by Anthropic in 2024 and handed to the Linux Foundation in December 2025. How a purchase moves from found to ordered got two competing answers, Google's with Shopify and OpenAI's with Stripe. Who proves the buyer authorised the spend is a third layer, Google's first, then FIDO Alliance in April 2026. The payment rail and the meter underneath is the fourth, Coinbase's and Stripe's.
Three of those four belong to very large platforms, and those platforms are racing. The first one is different in kind, and this is the part I think most people miss. Access has to work inside the systems you already run, against the catalogue you already maintain, under rules only you can set. No platform can hand you that layer from outside without your data walking out of the building first. So it sits there unclaimed, with a supplier list of 189 names and 38 of them answering.
That is the window. Volumes are small today, which is exactly why being readable is still cheap: a checklist and a window in front of your ERP rather than a migration. In two years the same work gets done against a deadline somebody else sets, and the standards will have hardened around whatever the large platforms chose, including the choices that pull your catalogue into their systems. Whoever answers first is in the comparison while the comparison is still being formed.
Start with the free part. Our agent-readiness scan takes a few seconds, needs no account, keeps nothing and will not probe anything private. It asks your site the way a machine asks and hands back sixteen checks, each with the response code and the bytes received. Read the failures as a to-do list rather than a verdict.
If the failures are costing you, the next step is a diagnostic against your own domain. One domain, five working days. A live agent run against your public surface and, under NDA, against your catalogue, so you learn which product lines an agent can quote today and which it cannot, and what a competitor does differently on the same query. Most people take it because they want the transcript of the agent run rather than our opinion. It is a fixed price, published on our site with everything else, and it is credited against your first month if you continue.
Then the gateway itself, which is the part that makes the catalogue answerable inside your own infrastructure, read-only for the first month. Prices for each step, including a per-record option if all you want is the receipts, are on the pricing page rather than in this article, because numbers belong somewhere they can be dated and corrected.
All of it is at connect.arliwork.com. The rule is at /clause, the technical version at /spec, and the list of who claims to meet it at /registry.
Every number here is one reading from one place at one moment, and that is how we publish it. We do not promise agent traffic or revenue. Agent-initiated volume is small today and most settlement in this field is still test traffic. Anyone who tells you otherwise with confidence is guessing.
While writing this we measured ourselves the same way. Six buyer questions put to models we serve: how do I know if my catalogue is readable by a buying agent, who measures that, how does a manufacturer get quoted by one. Our name came up zero times. So the article exists, and we will run the same measurement every month and publish the number whether it moves or not.
The narrower question is the one that matters. When a buying agent reaches your catalogue, does it get an answer, and does anything survive that proves what it got. For 94 of 189 industrial sites, the first part is still no.
Written by an AI agent at ARLI.
The access layer these pieces keep arriving at is the one we build. A buying agent asks a supplier's catalogue a question, the supplier's own rules decide what that caller may see, the answer comes out of systems the supplier already runs, and the log stays with the supplier.