What we will cover
A homeowner asks a browser agent to find a roofing contractor, check whether the company serves the address, and request a callback about a leak. The page looks fine. The service area is vague, the form label is missing, and the submit button moves when a chat box loads. The request either fails or reaches the office without enough detail to route it.
Agent readiness for a contractor is whether that request survives the handoff. It is not a race to install a new protocol across the whole site. It is a small operating test: can one customer task move through accurate service facts, usable controls, a clear confirmation, and a human owned follow up without changing scope, price, schedule, or safety promises?
Chrome's current guidance points to the same order of work. Agentic browsing checks and WebMCP are experimental. Semantic HTML, stable layouts, clear labels, and honest service pages already help customers. Build those basics first, then test whether an agent feature removes a real point of friction.
Choose one useful task, not the whole website
Pick a task that happens often enough to matter and is narrow enough to inspect. A good first test might be checking service area fit, choosing a service, or sending a callback request. Do not start with an emergency dispatch, a firm price, a signed change order, a payment, or a confirmed appointment. Those actions need stronger identity, authorization, business rules, and recovery controls.
Write the task in the customer's words. For example: find out whether this electrician serves my town and ask the office to call about a panel problem. That sentence gives the web owner a page to open, facts to verify, a form to test, and an office outcome to trace. It also exposes the limits. The site can accept a request. It cannot diagnose the panel, promise arrival, quote the job, or decide that the situation is safe.
Keep the first test reversible
A callback request can be reviewed, corrected, or declined by the office. A booked crew, accepted price, or changed schedule can create a customer promise before a person checks the facts.
Write the pass conditions before you run an audit
A tool can tell you that a control has a name or that the layout moved. It cannot decide whether the service area is real, the form reaches the right person, or the confirmation matches office practice. Put the operating pass beside the technical pass.
| Task stage | Customer pass | Machine pass | Owner and hold point |
|---|---|---|---|
| Find the service | The page names the work, property fit, area, and limits in plain text. | The title, headings, links, and page structure identify the same service. | Service manager. Hold when the page promises work the crew does not offer. |
| Choose the action | The callback or estimate action is visible and says what it does. | A real link or button has a clear programmatic name and state. | Web owner. Hold when a styled card or icon is the only control. |
| Enter the request | Visible labels and instructions explain the useful fields and formats. | Labels connect to controls and errors identify the field that needs repair. | Web owner. Hold when placeholder text or layout supplies the only meaning. |
| Confirm the change | The customer reviews what will be sent and chooses to submit it. | The action is marked as consequential and does not fire during discovery. | Business owner. Hold when the request can be sent without clear confirmation. |
| Reach the office | The page states the next step without promising unsupported timing. | The test creates one traceable request and returns a safe result. | Dispatcher. Hold when the alert is lost, duplicated, or sent to the wrong queue. |
| Close the loop | A person checks fit, urgency, scope, price, and scheduling. | The agent stops at the approved boundary and records a support ID on failure. | Assigned callback owner. Hold when software makes a field or commercial decision. |
Use one named owner for every hold point. The web developer owns markup and layout. The service manager owns service scope. The dispatcher owns routing and response practice. The business owner or an assigned manager approves any rule that could create a price, schedule, safety, warranty, or legal commitment.
Check the page the way a browser agent receives it
Agents may use a screenshot, the page HTML, or the accessibility tree. The accessibility tree reduces the page to roles, names, and states. If a visual card behaves like a button but is coded as a plain container, a person may guess that it is clickable while software misses it. If a field uses placeholder text instead of a proper label, the purpose can disappear after someone starts typing.
Google, part of Alphabet (GOOGL), recommends semantic controls, correctly connected labels, and stable layouts in its agent friendly website guidance. The World Wide Web Consortium gives the same label advice for people using assistive technology. The overlap is practical: the first agent readiness repairs are often ordinary accessibility and usability repairs.
- Use real links for navigation and real buttons for actions.
- Give every form control a visible, specific label connected in the code.
- State required formats, optional fields, and what happens after submission.
- Keep important service facts in page text rather than only inside an image.
- Reserve image and widget space so the form does not jump while it loads.
- Show a useful error near the field and preserve safe information the customer already entered.
Read the Lighthouse result as a developer signal
Chrome's experimental Agentic Browsing category does not issue the familiar score from zero to one hundred. It reports a fraction of passed checks and individual pass or fail results. The current checks include WebMCP registration, accessibility tree quality, names and labels, visibility, layout stability, and the presence of an llms.txt file.
That fraction is not an SEO grade, an AI recommendation score, or evidence that a service request reached the office. Google Search Central also says no special AI text file, machine readable file, or schema is required for inclusion in AI Overviews or AI Mode. Keep search visibility, browser task testing, and lead operations as separate questions.
Planning framework, not a benchmark: a limited browser agent test belongs in the upper right only after the human path works and state changes stay controlled.
A page can pass a technical check and still be wrong for the business. It can also fail an experimental WebMCP check while the normal customer path remains clear and usable. Record the exact failure, fix what affects people or the selected task, and leave speculative work out of the first repair list.
Protect the moment when the website changes state
Reading a service page is different from sending customer data, reserving time, accepting terms, or creating a work record. Chrome's current WebMCP security guidance treats actions with meaningful consequences differently from read only tools and recommends a confirmation step for high stakes actions. It also warns that prompt injection cannot be guaranteed away inside a language model.
For a contractor, the boundary should be visible in the task design. An agent may help the customer assemble a service request. The customer should see what will be sent and confirm it. The office should still decide service fit, urgency, price, availability, crew assignment, and any safety instruction. Do not expose a live dispatch board, customer record, payment action, or schedule change just because a demo can call it.
Log enough to investigate a failed task
Keep a safe record of the page, task version, submission time, result, and support ID without storing secrets or unnecessary customer details in public logs. A failed request should show the customer what happened and what to do next. The office needs a private diagnostic trail that can connect the support ID to the technical cause.
Add WebMCP only when it improves a working path
WebMCP is a proposed web standard that lets a page describe tools and form actions to browser agents. Chrome presents it as a progressive enhancement and currently offers it through an origin trial. That means the normal page still has to work. A contractor should not need WebMCP to explain services, show job proof, state the service area, or receive a basic request.
A limited WebMCP test may be useful when a form has structured choices that agents regularly misread, such as service type, property type, and preferred callback window. Define the expected inputs, keep the visible form and confirmation aligned, and test many plain language requests. Chrome's evaluation guidance notes that model behavior is probabilistic, so one successful demo does not establish reliability.
Stop the test if the agent chooses the wrong service, hides a required condition, submits without clear customer confirmation, or creates more cleanup for the office. The protocol is optional. The customer handoff is not.
Run a contractor field test in under an hour
Use a phone, a desktop browser, and someone from the office. Start from the service page a homeowner would actually find. Test the ordinary customer path first, then any agent assisted path available in your development environment. Do not use real customer data.
Set the task
Choose one service and area question with one safe customer action, such as a callback request.
Check the page
Confirm the service facts, proof, labels, errors, layout, and contact expectation with the people who own them.
Trace the request
Submit test information, inspect the confirmation, follow the alert into the office, and verify the named owner receives it.
Record the gap
Fix the first failure that blocks the task, then rerun the same scenario before adding another feature.
Use at least three variations: a clear fit, a location outside the service area, and a request with missing information. Add a safety sensitive example only if the responsible manager has approved the wording and escalation path. A good result is not that the agent finishes every request. A good result is that it completes supported tasks and stops cleanly when a person must decide.

GEO Smith helps connect a buyer question to the service facts, job proof, local signal, and human contact path that the contractor can support.
GEO Smith turns buyer questions into focused contractor visibility work.
See where AI style answers miss or misdescribe the business, then connect the useful gaps to approved service facts, local proof, job photos, citations, and a contact path the office can handle. Rerun the scan after the repair instead of treating one browser test as a marketing result.
See GEO SmithBuild the supporting path around the task
A browser agent task sits inside the same path a homeowner uses. These GangBoxAI guides help the team strengthen each part without turning an experimental protocol into a promise of rankings or leads.
- Use the broader agent ready contractor website guide to check service facts, proof, page structure, and approval boundaries.
- Tighten the lead form, confirmation, office route, and named callback owner before testing agent submission.
- Build the service page from current scope, coverage, limits, proof, and a next step the office supports.
- Keep page markup aligned with visible facts after the service and contact path are correct.
- Follow an AI referred buyer from the answer to a clear service page action and human response.
- Give changing services, areas, hours, and policies a source and an update owner.
- Use the GangBoxAI diagnostic when the team needs to choose the workflow problem before the tool.
- Use GEO Smith to find missed buyer questions and prioritize the proof and visibility work behind the task.
Sources
This guide uses current official documentation for the experimental browser features, website accessibility, and Google Search limits. Check the Chrome status and origin trial documentation again before planning production WebMCP work.
- Chrome for Developers: Lighthouse agentic browsing scoring
- Chrome for Developers: WebMCP
- Chrome for Developers: WebMCP tool security
- Chrome for Developers: Test WebMCP with evaluations
- web.dev: Build agent friendly websites
- W3C Web Accessibility Initiative: Labeling controls
- Google Search Central: AI features and your website
