AI Agent Protocol Adoption Study 2026: 287 Companies, 21 Protocols

Almost every number you have read about agent protocol adoption came from a survey or a registry count. Companies describing themselves, or a directory counting its own listings.

We measured the web instead. 287 companies, 21 agent protocol endpoints requested from each, every response validated. Then a second pass to see what each site does when a real AI agent shows up.

Five companies expose something an agent can call.

What we measured

Three populations: the Forbes Cloud 100, the 100 largest global retailers by revenue, and the largest US companies by revenue. Twelve companies sit on two of those lists, so each is assigned to one population and counted once. That gives 100 SaaS, 100 retailers, and 87 Fortune companies.

Every domain got the same treatment. Canonical origin resolved first, two random control paths to catch sites that return 200 for anything, then 21 protocol requests spaced 220ms apart. Our crawler identified itself and was never disguised as a browser.

One distinction organizes everything below.

A passive manifest is a document an agent reads. A file like llms.txt describes your site. It takes an afternoon and changes nothing about your infrastructure.

An actionable protocol is an interface an agent invokes. An MCP server, an agent card, a tool catalog. This means exposing real capability to an untrusted automated caller, which requires authentication, rate limiting, and an abuse model.

Every keynote of the past eighteen months has been about the second kind.

How many companies have adopted AI agent protocols?

Survey research cannot see this. When a study reports that most enterprise AI teams have MCP in production, it is measuring internal use: an MCP client wired into a private developer tool, pointed at somebody else's server. Both numbers can be true at once. Companies are enthusiastic consumers of other people's agent interfaces and almost universally unwilling to publish their own.

Which AI agent protocols have zero adoption?

llms.txt is the most adopted protocol by a wide margin, and it splits hard by sector: 54% of the Cloud 100, 14% of the Fortune 100, 6% of the largest retailers.

Below that, the actionable rows are not small numbers. They are zeros.

Agent2Agent is the clearest case. It reached version 1.0 under the Linux Foundation with more than a hundred organizations behind it, a governance model, and a conformance suite. We requested its well-known path from all 287 companies, including several on its own partner lists, and found nothing.

The fair reading is that A2A adoption is real but private, running between contracted partners at authenticated addresses nobody publishes. That is a legitimate architecture. It also means the public web offers agents no discoverable front door. This is a narrower claim than the launch partner count, which measures endorsement and in-product support rather than a published, callable endpoint, and our probe checked each company's primary domain rather than app subdomains or per-agent paths.

Why this is a demand problem, not a deployment problem

The obvious objection is that large organizations simply cannot ship static JSON at a fixed path. CDNs, change control, committees.

So we measured security.txt as a control. Same directory, same deployment process, same review boards, no relationship to AI.

Companies adopted security.txt because a norm formed, ignoring it carried a reputational cost, and the benefit was legible to whoever approved the ticket. Agent protocols have none of those conditions yet.

Schema.org adoption across 287 major brands

Structured data is the oldest and best supported way to make a page machine readable. We recorded every JSON-LD type and every microdata itemtype on every homepage we could retrieve.

An important denominator note: 184 of the 259 responding companies returned a homepage to our identified crawler. The rest refused it. Every percentage below uses those 184, not the full list, because a site we could not fetch is not evidence of absent markup.

The one action markup companies do implement

Schema.org has types that describe and types that let something be done. Here is the count across 184 homepages.

SearchAction appears on 42 homepages with its companion EntryPoint on 30. It is the only action capable markup with meaningful adoption, and it exists because Google rewarded it with a visible search box.

OrderAction, BuyAction, and ReserveAction are the vocabulary schema.org provides for exactly the transactions an agent would want to complete. Not one company declares any of them. Speakable has been recommended for years and is easier to implement than SearchAction. Two companies use it.

One caveat, because it is the first fair objection. We measure homepages, and Product and Offer markup normally lives on product detail pages, which is why those counts are low here at 7 and 8. That does not rescue the finding. OrderAction and ReserveAction belong on a homepage and remain at zero.

The pattern is not difficulty. The thing nobody implements is the easier one. Companies ship machine-actionable markup when a platform pays for it in traffic or placement, and not otherwise.

That is the most predictive finding in this dataset. Agent protocol adoption will not move because a standards body ships a spec. It will move when a major assistant starts routing real transactions to sites that expose a callable interface.

What the best case looks like

One company in 259 scored an A. Forter, a fraud prevention firm at Cloud 100 rank 65, serves an MCP discovery document, a populated agent card, and a live natural language endpoint. A six protocol outlier is exactly the shape of a measurement error, so we verified it by hand against control paths.

https://www.forter.com/.well-known/mcp.json

A complete MCP discovery document, with a caveat that matters. The availability status reads demo-only. This public server does not call into real Forter risk decisioning. The production server lives at mcp.forter.com behind a signed master service agreement.

Even the best result in the study is a sandbox. Its real agent interface sits behind a contract and a sales conversation.

https://stripe.com/llms.txt

The common case. A well maintained llms.txt pointing agents at documentation. Useful, and entirely passive.

https://www.walmart.com/.well-known/agent-card.json

The world's largest retailer, asked to identify itself to an agent. The response is a 404 page whose Content-Security-Policy enumerates several hundred advertising and analytics domains.

Nineteen companies return HTTP 200 for paths that do not exist. A study trusting status codes alone would have credited all nineteen with adopting every protocol we tested. That is one concrete way inflated adoption numbers get made.

The full ranking

The Citerank Agent Readiness Score weights actionable protocols at 35 points, discoverability including crawler policy at 32, semantic structure at 20, and commerce manifests at 13. Every point traces to an observed artifact, so the score is recomputable from the published dataset.

What happens when an AI agent actually shows up

Everything above measures what companies publish. It does not measure what happens when an agent arrives.

So we requested the same homepage four times from the same address, changing only the identity: Chrome, then GPTBot, ClaudeBot, and PerplexityBot. A site that serves the browser and refuses the agent is running a gate that robots.txt never mentions.

This is the number no survey or registry can produce. The common story is that publishers are barricading themselves against AI crawlers through robots.txt. Almost nobody writes the rule. A quarter enforce one anyway, through a bot management layer that never announces itself.

For an agent, that is worse than an honest block. A robots.txt rule is a stated position it can respect and route around. A silent 403 from a security vendor looks like an outage.

Can an agent complete a task on these sites?

Getting served is step one. An agent sent to buy something or compare plans has to find the relevant page from what the server returns.

There is a compounding problem underneath. On 17 of 74 retail sites the homepage arrives as an application shell with almost no readable text, because the content is assembled in the browser afterward.

Stack the three failures and retail looks like this. Four in ten refuse the agent. Of those that let it in, a fifth return a page with nothing in it. Of the pages with content, most carry no pricing link.

Three industries, three different failures

The Cloud 100 builds the agent economy and holds all five actionable implementations. It also shows the gap most sharply, since 54% publish llms.txt and 5% expose anything callable. Adoption is flat across the list, so this is not a resource question.

The largest retailers have the most to gain from agentic commerce and the least to show. Median readiness score of 10. Five of 89 publish llms.txt. Half of the homepages we could retrieve carry no structured data.

The Fortune 100 sits between them. It publishes llms.txt at 14% and security.txt at 24%, which tells you these organizations ship well-known files at the pace of any other compliance artifact. Nothing in the agent stack has become one yet.

Does budget or market position predict agent readiness?

The comfortable explanation is resources. The data does not support it.

Rank is a crude proxy, so we tested capital directly. The Forbes list publishes total funding raised for every scored Cloud 100 company.

Headcount tells the same story at -0.07. What separates a prepared site from an unprepared one is not money, size, or market position.

Checking our score against an independent scanner

Ora, an agent readiness company that partnered with Vercel to launch is-agentic.com, runs its own scanner and reports scanning more than 59,000 sites. Their tool grades an individual site and publishes results through a public API.

Their work answers a different question. They score sites one at a time, largely ones submitted to them. We measured three defined populations end to end, which is what allows a statement like "zero of the hundred largest retailers" to mean anything. The overlap gave us a way to check ourselves.

Their scoring only counts checks that apply to a given site, so a retailer with no developer platform is not marked down for lacking an OpenAPI description. Ours does mark it down, because the question here is not whether a site is well built for what it chose to offer, but whether an agent arriving cold can do anything with it.

Both framings are defensible. What matters is that a second, independently built scanner puts these companies in substantially the same order.

How this can be wrong

One correction during the study is worth publishing, because it is a trap for anyone repeating this work. We disabled redirect following so a 301 to a homepage could never count as a hit, and the first implementation then read every apex domain that redirects to www as barren. Cloudflare went from zero detected protocols to nine once canonical origin resolution ran first.

A second correction changed the schema results materially. Our first pass counted sites whose homepage never returned 200 as having zero structured data, and parsed only JSON-LD. Both were wrong. Excluding unfetchable homepages and adding microdata moved structured data adoption from an apparent 38% to an actual 66%.

The known limits. 28 companies never responded and are excluded rather than zeroed. A further 75 responded to protocol probes but refused a homepage request, so schema figures use a smaller denominator, stated everywhere it applies. We probe the apex domain only, so a company running an MCP server on a documentation subdomain is not counted. Findings describe public web surfaces on the crawl date and say nothing about private or contracted integrations, which is where the A2A traffic appears to be.

Our score also penalizes a site for lacking a callable interface even when it has no developer platform to expose. That is a choice, not an oversight. A retailer with no public API is not badly engineered. It simply cannot be transacted with by an agent, and that is the thing this study counts.

What this means for you

If you sell to businesses, the discovery layer is cheap and most competitors skipped it. A valid llms.txt, clean structured data, and a deliberate robots.txt position put you ahead of roughly half your market. Our guide to building agent ready websites covers the implementation order.

If you are in retail, the current bar in your sector is zero. A retailer shipping a working agent interface would be the only one, at the moment assistants are looking for somewhere to send purchase intent.

If you have never measured any of this, that is the normal state. 147 companies in this study returned nothing at all. Our earlier study, where we scored 19 SaaS brands and none passed, found the same pattern in miniature.

The question this opens

The full dataset is published as a CSV covering all 287 companies, with one column per protocol so any figure here can be recomputed.

This study measures supply. It does not measure whether any of it works.

The obvious next question is the one the field argues about constantly: does publishing llms.txt change how often a brand gets cited by AI systems? This dataset produces a usable natural experiment, with 54 adopters against 46 non-adopters inside a single industry list, measured on the same day.

We are running that next, controlling for funding, authority, and brand demand so the answer survives the objection that large companies both publish llms.txt and get cited anyway. The free AI Visibility Score will tell you where your own site sits in the meantime.