← EYMA Dispatch

What AI Agents Actually Read on Your Website Before Contacting You (and What Makes Them Leave)

EYMA · Published November 11, 2026 · Updated November 11, 2026

AI agents do not browse a website the way a person does. They are not reading your hero copy, admiring your brand photos, or following your navigation. They are scanning for structured signals that answer a narrow set of questions: What does this business do? Where does it operate? Is it licensed and verifiable? How does it accept agent contacts? When those signals are present, the agent can proceed. When they are absent, most agents record the business as low-confidence and move on.

Understanding this changes how you think about your site's job. For human visitors, a business website is a persuasion instrument. For an AI agent, it is a data extraction target. The question is not whether your copy is compelling. The question is whether your site's machine-readable layer gives an agent enough to work with before it decides whether to contact you.

The two layers are not in conflict. A site can be optimized for human readers and machine readers simultaneously. But most businesses have only ever thought about the human layer, which means the machine-readable layer is often absent entirely. That gap is what this article addresses.

What parts of a business website does an AI agent actually parse?

An AI agent parses the structured elements of a website built for machine reading: page title, meta description, header tags, schema markup, and any machine-readable files the domain publishes such as registry.json or llms.txt. It does not read a website the way a person does. It extracts the signals it was trained to find, including business type, service area, license status, and contact path, then moves on. If those signals are absent or buried in unstructured text, the agent records them as unknown.

The most reliable signals are those delivered through schema markup. A LocalBusiness schema block on your homepage communicates your business name, type, geographic area, and contact information in a format any agent can parse without interpreting prose. If you operate in a licensed field, a schema block that includes your license number is worth more than a paragraph on your About page that mentions it. The structured version is queryable. The prose version requires inference.

Machine-readable files at the domain root are the second tier. The EYMA registry feed at eyma.ai/registry.json is an example of this layer: a file structured for agent consumption, not human reading. The intake specification template published at eyma.ai/how-agent-to-agent-trust-works describes how businesses can expose a machine-readable contact spec alongside their existing site, so incoming agents know exactly how to reach a human reviewer, what data to include, and what format the business accepts. This is the difference between an agent guessing how to contact you and an agent following a published protocol.

The third tier is the content layer: header tags, page titles, and meta descriptions that mention service categories, geographic coverage, and license information in plain language. These are less precise than schema markup but still better than nothing. An agent inferring from an H1 that says "Commercial Insurance in Sacramento" has more to work with than an agent reading a homepage headline that says "Protecting What Matters Most." Descriptive headers serve both human and machine readers.

What signals tell an AI agent a business is worth contacting?

The signals that clear an AI agent's threshold for contact are clarity of service category, verifiable identity, and a defined intake path. A business that names what it does, publishes a license reference tied to a government database entry, and provides a structured intake channel is worth contacting. A business that requires the agent to infer all three from prose gives the agent less to work with and more room to reach the wrong conclusion.

Verifiable identity is the signal that is hardest to supply through the website alone. A license number in a schema block tells an agent the number. It does not tell the agent whether the number is current. For that, the agent needs either a live government database query or a trust registry entry where someone has already done the verification work. This is why an external listing in a registry like EYMA complements website signals rather than substituting for them. The website communicates what the business does and how to reach it. The registry entry confirms the business is what it claims to be.

The intake path signal is the most practically important. An agent that has identified a business as a credible match for its principal's needs still needs to know how to contact it in a way the business will respond to. A contact form that does not distinguish agent contacts from human inquiries is not an intake path. It is a queue where the agent's structured request will sit alongside a retiree's question about coverage, waiting for a staff member to notice the difference. A published intake specification, by contrast, tells the agent what to send, where to send it, and what response to expect. Setting up an email-only agent intake is the minimum viable version of this, and it is achievable without custom development.

What makes an AI agent abandon a website without making contact?

An AI agent leaves without contacting a business when it cannot answer the minimum required questions from the site's structured layer. If the business type is ambiguous, the service area is unspecified, there is no license reference, and there is no structured intake path, the agent has nothing to anchor a confident contact decision to. It may flag the business for a human review in its principal's workflow, or it may simply move to the next candidate. Either way, the business does not receive the inquiry.

The abandonment is invisible. The agent does not send a failure notice. The business never sees a "could not verify" message. What it sees is an absence: no inquiry, no quote request, no contact attempt. For a business that has not yet seen agent-originated traffic, this is indistinguishable from normal operations. For a business that is expecting agent traffic in a category where agent commerce is active, the absence is the signal.

The agent-readable business profile standard describes the full picture of what a machine-readable business identity looks like across all surfaces: website, registry entry, intake spec, and license reference. Building all four is not a large project. The website changes are schema markup and a few meta updates. The registry listing is a form. The intake spec is a published template. The license reference is a copy of what is already in a government database. What makes the combination powerful is that it removes ambiguity from every layer an agent touches. EYMA is the place where legitimate licensed bots go to sell their humans' products, and the businesses that get the most agent traffic are the ones that removed every reason for an agent to leave before making contact.

The mechanics of how agents find local businesses and the cost of an unregistered agent identity describe the two sides of this problem: how agents discover candidates and what a business loses when it is not in the right form to be found. The website is the discovery layer. The structured trust record is what converts discovery into a contact. Neither works well without the other.