Your website may already be answering AI agents. You didn’t write what it says.

What was measured

Public HTTP GETs of the documents an AI agent looks for — /llms.txt, three/.well-known/ documents, robots.txt, a sitemap, and the organisation data on the homepage — following each site’s own robots.txt. Nothing was submitted. No page was altered. No business was contacted.

Businesses were found in OpenStreetMap’s open data, filtered to those publishing a website of their own, across 10 areas. Two rules govern every number here. A check that could not complete is recorded as unknown, never as absent — a server that answered 503 is a server that failed to answer, not a business that publishes nothing.

And no percentage on this page is a percentage of 6,284. Each share is taken over the businesses where that check could be decided, and that sample is different for each one. For /llms.txt it is 4,856: The study covers 6,284 businesses. 4,856 had an assessable /llms.txt status; the remaining 1,428 were not assessable for this measure and are excluded from it rather than counted as publishing nothing. Every table below prints its own assessable count in the column beside the share, so the denominator never has to be inferred.

1,887 businesses could not be read well enough to assess overall. That figure is the study’s own coverage limit and is not the denominator of any percentage here — each check has its own, in the table above.

1. Presence arrived by default

A year ago /llms.txt was a proposal. Today 1,242 of the 4,856 businesses where we could decide the question publish one.

That looks like an adoption curve. It is a release note.

Of the 1,238 files we could re-read, 734 name their own author — and the authors are five pieces of software.

The other 504 name no author. We count that as no author named — not as “written by the business”. Plenty of software does not sign its work.

And it goes past a text file. 498 of these sites declare a working agent interface, in as many words: “This site is powered by Wix and supports the Model Context Protocol (MCP) for agentic AI access.” A garden centre in Alabama and an ice cream maker in Tennessee are both queryable by an assistant today, over a live protocol, because their storefront platform shipped it. Nobody at either business decided that.

2. Presence is not control

Here is what none of them publish — and “none” means none of the 4,385 businesses where the question could be decided, not none of the 6,284. On the other 1,899 we could not read far enough to say either way.

A platform-generated file says what the platform decided to say. It is assembled from the pages and products already on the site. It does not say what the business will quote a price on and what it will not. It does not say which services it actually performs versus which it lists for search traffic. It does not say the hours it will honour, the area it will travel to, or which questions it wants a human on.

An agent reading a platform-generated file learns the site’s page structure. An agent reading nothing at all learns the same thing from the HTML. In both cases the business is the subject of the description rather than its author. That is the gap. Presence is now free and automatic; accuracy, capability and answering on your own terms are not.

3. Three quarters are outside even the automatic layer

3,614 of 4,856 businesses — 74% — publish no /llms.txt at all. These are disproportionately the businesses not on a platform that ships one: independent sites, older builds, custom work, and small firms whose site was built once and left alone. No plugin update is coming for them.

A further 213 block AI crawlers outright.

We make no claim about whether blocking is right or wrong; some businesses have good reasons. What we observe is that it is almost always all of them at once, which is what a pasted line looks like rather than five separate decisions.

4. What we got wrong while measuring this

Two findings were caught before publication. A third was not, and it was published about named companies before we caught it.

We have withdrawn the organisation-data column, and we had it wrong about businesses we named. Until 2026-09-03 this study reported organisation data as present on 2,012 of the 4,749 businesses whose homepage could be read for it (42%), absent on 2,737 of them, and not readable on 1,535. Every one of those 2,737 businesses was named on this page and in the dataset. The check behind this column reports absent for homepages that publish Organization and LocalBusiness data — the plain type names, not only the narrower subtypes we already knew it missed. On 2026-09-03 we re-fetched a fixed-stride sample of the homepages this study had called absent and read the raw HTML ourselves: 53 of 196 readable domains, and 52 of 196 on a second run the same day, publish a business type we had recorded as missing. That is a floor of at least 26.5% of the rows we called absent — roughly 726 of the 2,737 — and a floor only, because a type the audit does not recognise is reported rather than counted. We confirmed it against the live checker the same day, so this is our measurement being wrong and not the sites having changed.

So the column is gone rather than corrected. It reads withdrawn in the table above and on every one of the 6,284 rows in the dataset, including the ones it previously called present. We are not publishing a replacement percentage: we know the old number was wrong and we do not know the right one, and a figure we cannot derive is exactly the kind of thing we would criticise someone else for printing. The same signal has been removed from the comparison on our free scan, which no longer measures anyone against it. When the check is fixed, the sites will be re-read before anything is republished — a fixed checker does not retroactively correct a reading taken on 2026-08-31.

If your business is in this study, this is the line that was wrong about you, and it was wrong in the direction that flattered our argument. We would rather say so here than quietly delete it. Tell us if anything else is wrong and we will do the same again.

We built a scoring band only reachable by buying our own product. Our assessor bands a site “agent-ready” at 70 out of 100. The highest score any business achieved was 57. Three of the seven checks are the agent-presence documents this firm publishes and sells; they were present zero times, so the top band was arithmetically unreachable. “No small business is agent-ready” would have been true against our own definition and false to every reader. No business is given a score anywhere on this page or in the dataset.

We reported 498 businesses as having no agent interface, because we only looked for our own filename. The check reads /.well-known/agent-interaction.json and nothing else, while Wix was writing MCP support into files we had already downloaded. A measure that can only see its author’s spec reports the rest of the world as empty. The figure above is a floor, not a total: MCP was only looked for inside llms.txt files.

We publish all three because a study that reports only its successes is asking to be trusted rather than checked — and the third one had already gone out under our name before we found it.

Look up any business in the study

Every one of the 6,284 businesses is listed with the facts observed about it: which documents were present, which absent, which could not be read, and which we have withdrawn. No business is scored, graded or ranked. We report what a URL returned. What that is worth is your judgment, and theirs.

Or read them as pages: the 4,856 businesses whose /llms.txt we could decide have one each, carrying the address we requested and the HTTP status it returned — New York City, Alabama, Chicago, Philadelphia, Austin, Atlanta, New Orleans, Nashville, Charlotte, Memphis.

Download the complete dataset — 6,284 businesses, free to reuse with attribution — as CSV or as JSON. The JSON carries a machine-readable corrections block, so a copy fetched before 2026-09-03 can learn what we withdrew.

The ten areas, one page each

The national figure is a fact about a sample most readers are not in. These are the same checks, taken the same day, narrowed to the place a business actually operates in — with the bounding box that was crawled printed on each page, and every business in it named.

• New York City 1,494 read · 828 of 1,106 assessable publish nothing for agents (75%)
• Alabama 984 read · 596 of 762 assessable publish nothing for agents (78%)
• Chicago 890 read · 499 of 695 assessable publish nothing for agents (72%)
• Philadelphia 676 read · 410 of 551 assessable publish nothing for agents (74%)
• Austin 670 read · 394 of 533 assessable publish nothing for agents (74%)
• Atlanta 535 read · 295 of 408 assessable publish nothing for agents (72%)
• New Orleans 343 read · 191 of 245 assessable publish nothing for agents (78%)
• Nashville 300 read · 172 of 242 assessable publish nothing for agents (71%)
• Charlotte 286 read · 165 of 229 assessable publish nothing for agents (72%)
• Memphis 106 read · 64 of 85 assessable publish nothing for agents (75%)

What this study does not claim

• Nothing about outcomes. We measured what is published, not whether it changes how often a business is surfaced, cited or recommended by any assistant. That is a different study and we have not done it.
• Nothing about the 504 unattributed files. They may have been authored, or generated by software that does not sign its work. We do not know which.
• Nothing about the 1,887 businesses whose documents could not be read, beyond the individual checks that did complete.
• Nothing about businesses without a website in OpenStreetMap — a different and almost certainly larger population than this one.
• Nothing about whether blocking AI crawlers is a mistake. We counted it.

Check it yourself

Every figure on this page is generated from the dataset rather than typed, and a verification gate refuses the build if any of them stops reproducing. Any business listed can rerun the same checks on its own domain in a browser.

This is a point-in-time reading. One business in our sample was recorded as having no /llms.txt and returned a 44KB one when re-checked an hour later; we do not know whether the site changed or our earlier read differed, and we have not named it as either. Any business that wants to disagree with its line can change the answer today.

If we have something wrong, we would like to be told, and we will correct it in place with a note saying what changed. Tell us, or check your own site free.

Study by River Cade Concepts. The scheme behind the check is published in full.

Cite this study

River Cade Concepts, Agent presence among 6,284 small businesses, 2026-08-31, corrected 2026-09-03. https://rivercadeconcepts.com/reports/agent-presence

The data is free to reuse with attribution, and we would rather you re-derived a figure than trusted ours. One request, and it is the same one we failed ourselves: if you quote a percentage, quote the sample it was taken over with it. Every share here is over the businesses where that particular check could be decided, and those samples differ by thousands. A number from this study that arrives without its denominator has already lost the thing that made it true.