Greg Isenberg's bet is that billions of dollars are about to move through AI agents, and the internet isn't built for them yet. Send an agent to get a quote on a site built for people and it has to work the way a person does. It reads the page, guesses which button to press, and fills in a form built for eyes.
WebMCP is meant to change that. Isenberg describes it as a standard from Microsoft and Google where, instead of the agent guessing, the site just tells it what it can do: here's how to search, here's how to book, here's how to buy.
Isenberg pitched two businesses around it: an agency that makes business websites agent-ready, and an agent mystery shopper that tests whether agents can finish the jobs that matter. Both work just as well pointed at your own site.
List the jobs customers come to your site to finish
Two or three, no more. Request a quote. Book a consultation. Buy the thing. Check whether you do a particular job in their area. These are the jobs an AI agent will be sent to do on a customer's behalf.
Greg Isenberg's framing of why this matters: SEO was about Google understanding your page, AEO was about AI citing you, and WebMCP, a standard from Microsoft and Google, is about the agent actually finishing the job.
Checkpoint
Two or three jobs, each one a thing a customer finishes.
Mystery-shop your own site with an agent
the step that pays off
Give an AI assistant that can open and use web pages the prompt below, once per job. It stops short of submitting anything, so nothing real gets booked.
This is Isenberg's "agent mystery shopper" idea, turned on your own site: test whether an agent can complete the jobs that matter, then write down where it got stuck, which tools were missing, which descriptions were bad.
Is Agentic, in our directory, scores a site's readiness for agents and gives recommendations, which is a quick second opinion.
Send an agent to finish one job on your site
Pro tip
Run each job twice. An agent that finishes once by luck and fails the second time is still guessing.
Checkpoint
One report per job, with every guess the agent had to make.
Write the actions an agent should call directly
Isenberg's plainest line on WebMCP: "WebMCP is basically websites with agent buttons. (In the margin: the whole idea.)" Instead of an agent scanning a page like a human, taking screenshots and guessing where to click, the site tells it how to search, how to book, how to buy.
Those buttons need words. Paste your jobs, the mystery-shop reports and your business facts into the prompt below. The descriptions are the part to read closely: an agent picks an action by its description, and a vague one is the "bad descriptions" failure in Isenberg's list.
Write the spec for the actions an agent should call
Checkpoint
A spec with one action per job, each with a human step where money or a booking is involved.
Hand the spec to whoever builds your site
Give the spec to your developer or site builder and ask what it takes to offer those actions to agents on your platform. Isenberg's list of first tools for a business site is short: request a quote, book a consult. Start there even if the spec has more.
Keep the human step in whatever gets built. An agent can request the quote. A person on your team confirms it.
Checkpoint
The first one or two actions live, with a person confirming each result.
Mystery-shop again after the change
Run the same prompt against the same jobs. Compare the reports with the first ones: the steps, the guesses, whether it finished. Keep both reports. They're the before and after for this change and the baseline for the next one.
Checkpoint
Two reports per job, before and after.
The takeaway
You end up with a report on where an agent gets stuck on your site today, a spec for the actions it should be able to call instead, and a before and after once the first action is live.
The limit: WebMCP is new, and which agents use it, and how, will keep changing. The mystery shop is useful either way, since what stops an agent on your site often stops a person too. Confirming prices and bookings stays with a person. An agent can report a job finished that a person on your team would have caught. (In the margin: keep the human step.)
Variations
Fix the page before the protocol. Most of what the mystery shop finds, an unclear button, a form field with no hint, a price you have to hunt for, also slows down people. Fix those on the page first. Those fixes help every visitor.
Shop a competitor. Run the same prompt against a competitor's site for the same job. If their agent run finishes and yours doesn't, that's the case for doing this.
One job at a time. Build request a quote first, prove it with a second mystery shop, then add booking. A spec with ten actions and no tests is ten ways for an agent to go wrong.
Don't let an agent pay or confirm unattended. Payment and confirmed bookings go through the human step in the spec, whatever the agent on the other side is allowed to do.